Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources

[seathatflowsinourveins]

Scope

• What this PR changes, in one or two sentences: the follow-up to Ground the Claude Code native practice in primary sources and one cross-client mapping #958 for the four P3s of the command center's Claude micro at 7b216145 (the advisor-spend commit) and its two optional ones, which Ground the Claude Code native practice in primary sources and one cross-client mapping #958 landed without: the reading's run text, the pricing read time, the cache-split check and the assumed speed (four fields and tests of the advisor-spend record), and the Claude column of reference/cross-client.md pinned to documentation pages, recorded capture hashes and lines, each citation validated on its own.
• Base commit: 03a986ed637349403658f23da694c1833527819d (main with Pin the claude-native-practice skill as a first-party manifest row and raise the listing ceiling to 11,000 #972 landed); eight commits: e42246411cea59f3d1f8865e374b527b6b22af6d and 428fcb19108ab1a9f7a3b9b58e15d70cb17f7d02, replayed onto it from the follow-up prepared on Ground the Claude Code native practice in primary sources and one cross-client mapping #958's merge (no conflict), cc454f7f80373a39ba28cd35c9c9ca6e26cd2a90, the forward commit that answers the GPT read of 428fcb19, and f085a62b20e6add05dba0e56b497d0d0437ebffb, the one forward commit that answers the GPT micro at cc454f7f and carries the command center's two P3s of its read at that head (both below), and be54c82ef067cada6124c709ac8fe674fc1bb1aa, the one forward commit that answers the Codex review in GitHub and the co-op's GPT micro, both at f085a62b (below), and 18227acd3478656af8efb0d0023b67f76188069e, the one forward commit that answers the command center's Claude micro at be54c82ef (below), and 23fa8ef7a4589de6178156533c286f503dab2518, the one forward commit that answers the command center's Claude micro at 18227acd3 and Codex's two threads of the same head (below). The range for a read of that commit was 18227acd3..23fa8ef7a; the co-op's GPT micro reads f085a62b..23fa8ef7a. The eighth commit is a7dac4fdf4cc0da6dff62469d3cc0bfabbc76a6f, the final fix round (below): it answers the command center's Claude micro at 23fa8ef7a (three P2s) and its adjudication of 2026-10-11T13:50Z of Codex's two P1 threads of 13:35:12Z, and writes the tests' invisible characters as escapes; the range for the confirmation micro is 23fa8ef7a..a7dac4fdf.
• Lane: lane:foundation
• Owned paths touched (6 files): docs/decisions/2026-10-09-claude-code-native-practice.md is untouched; the changes are evidence/artifacts/claude-native-practice-20261009/primary-source-reading.json (usage.pricing read time and note, the speed assumptions, the run text), tests/test_primary_source_reading_counts.py (four tests, and two in the eighth commit; one is skipped unless PLATFORM_DOCS_CAPTURES names the directory with the retained advisor tool page), .claude/skills/claude-native-practice/reference/cross-client.md (a source table under the mechanism table), tests/test_native_practice_cross_client_sources.py (new, thirty-five tests; three are skipped unless CLAUDE_DOCS_CAPTURES names the captured pages (one) or GFM_RENDER_CHECK is set (two)), evidence/artifacts/claude-native-practice-20261009/cross-client-lines.json (new: the cited lines with their hashes) and manifests/evidence.json (re-registration).

The skill folder changes (reference/cross-client.md), so the manifest row of #972, which pins this skill at 7a6bfbcf0, stays at that pin until ONE re-pin PR that follows this PR, the PR that retires the Claude-in-GitHub rows and the PR that records the paired-run results, and names the three PRs' row sets (the command center's rulings of 2026-10-10 and 2026-10-11). No personal copy of the skill is installed on this host (python3 scripts/skills_status.py --json: canonical, link and lock missing), so sessions here read the project folder; the row is not to be installed before the re-pin. SKILL.md is unchanged.

Response to the GPT read of 428fcb19 (CHANGES_REQUESTED, one P2; full verdict ~/.cache/ns2604-gpt-reads/full1-pr986-428fcb191/review-output/VERDICT.md, 30,592 B, sha256 59ec9fecb2ecaae153856a83f1a2fc1767dfb3134f212e8f7f1f7d3f91349049). The first guard parsed a cell with findall and asked for one recognised citation, so a corrupted citation vanished whenever another citation of the cell stayed valid; its mutation test changed both Completion gate citations together. Reproduced here first: the first version passes 75 of the 154 single-citation corruptions of the page (the reader's four probes on the first hooks.md citation among them). The forward commit cc454f7f80373a39ba28cd35c9c9ca6e26cd2a90: each citation is split out of its cell and must match the whole citation form (malformed text is a violation); EXPECTED lists the (page, line) citations each need rests on, in order, so losing, gaining, reordering or moving one of several is caught; and the mutation test changes or removes each citation of each cell by itself (14 corruptions of each of the 11 citations). The fixed module passes none of the 154 mutants and still passes the unmodified page (probe table sha256 4e3a48f56a494b71f23236f58ef060712d13b5244d2255fb79cd1d8add31d8c1).

Response to the GPT micro at cc454f7f (CHANGES_REQUESTED, one P2 and one P3; full verdict ~/.cache/ns2604-gpt-reads/fix1-pr986-cc454f7f8/review-output/VERDICT.md, 33,426 B, sha256 a431713f07833974dd681bdb13d0f8973fe8df5a427f2004c2be029751aaa45f). The original P2's four reproductions are fixed and its 154 mutants are all rejected (the read replays them independently). Its remaining P2 is right and reproduced here before any change: the citation pattern still ended in a free note group (.+), which fullmatch let consume the note's closing parenthesis, trailing text and another note's opening parenthesis, so appending (extra prose), the first citation after a space or after ; see , the same on invalid.example, or the citation with no separator, each left violations() empty and the module green (five variants, 0 violations on the old guard). The note group is now [^()\[\]\n]+, so a note ends at its own closing parenthesis and holds no bracket. Five more corruptions are mutated into every citation (209 mutants, the 154 kept), the five probes are a test of their own and the note's boundary is asserted. On the old note group the new tests fail (62 failing subtests: 56 in the mutation test and 6 in the probe test); on the new one all six tests of the module pass. Its P3 asked for the upstream mechanism at a pin: the module's docstring now cites CPython 3.13.16 Lib/re/__init__.py lines 169-172 and the re documentation of that release, which say fullmatch applies the pattern to all of the string and do not bound what a greedy .+ consumes (so the boundary is in the pattern).

The command center's two P3s of its read at cc454f7f (PASS_WITH_P3), held until a finding of the other read allowed a push (its ruling of 2026-10-11). primary-source-reading.json: the run text says the refused advisor calls add no advisor charge (the measured part is that a refused call carries no usage: 34 too_many_requests results and no advisor iteration for any of them) and that the executor tokens that issue a refused call stay in the executors' spend, instead of "cost nothing", which no page line prices; usage.pricing records hashed_urls (the .md URLs whose bytes the recorded hashes are of, fetched by p958_p3_fix.py) next to source and advisor_billing_source (which name the HTML pages), and its read_note names the script and its sha256. Two assertions of tests/test_primary_source_reading_counts.py pin both and fail on the artifact without the change.

The command center's micro at f085a62b (PASS_WITH_P3, Gate line PASS). Its one P3 is fixed in this description and the head does not move: the description called the guard module "(new, five tests)" and one evidence-class row said 5 tests; the module has six (python3 -m unittest tests.test_native_practice_cross_client_sources prints Ran 6 tests and OK at f085a62b, re-run for that edit; the forward commit below makes it eleven).

Response to the Codex review in GitHub at f085a62b (P2, the page's Claude column sources paragraph) and to the co-op's GPT micro over cc454f7f..f085a62b (CHANGES_REQUESTED, one P2; full verdict ~/.cache/ns2604-gpt-reads/fix2-pr986-f085a62b2/review-output/VERDICT.md, 19,907 B, sha256 57b8e7b253b4b1c9facd255ef883c0426f7e50df33952f0c5c19e179c90addbf). Both are right, and both were reproduced here before any change. One forward commit, be54c82ef067cada6124c709ac8fe674fc1bb1aa, answers both.

• Codex: the page called the line of a citation a locator of claims.json. That file (blob cb0fe6c39497d0f5c6261cd97315d1117acc3a86) holds 257 documentation locators (256 with a capture hash and 253 with lines; this description first said 256, corrected below), but only 2 of the page's 11 citations (the two cli-reference lines, 103 and 112) fall inside a range recorded for the same capture. cross-client-lines.json now records the 10 distinct cited lines, each with the capture's full sha256 and fetch time (both agree with claims.json, checked by the guard) and the sha256 of the line's text (what sed -n '<line>p' <capture> | sha256sum prints for a line that is followed by a newline, which each cited line is). The page says exactly that. The captures are private working papers, so every cited byte is retained with a hash beside the other research records: the seven documentation pages (each cited line was read there and says what its cell says), the pricing and advisor-tool .md pages (67904e28..., 81a97b99..., unchanged), the Python re page (19195ec7..., the same bytes again), the writer script (d21b6dc2...) and the page of the new citation below. CLAUDE_DOCS_CAPTURES=<dir> python3 -m unittest tests.test_native_practice_cross_client_sources recomputes every recorded line hash from the bytes where they exist.
• GPT micro: a duplicate source row hid a malformed citation. page_rows() stored cells[need] = cell, so a later row replaced an earlier one before any check saw it. Reproduced on f085a62b with the reader's pages: a wrong host, trailing prose or an exact copy placed before the valid Completion gate row leaves all six tests green (a wrong host or trailing prose after the valid row, or in its place, fails; an exact copy passes in either position, because the replacement is the same text). The mechanism is Python's mapping assignment, documented as d[key] = value: "Set d[key] to value" (docs.python.org/3.13/library/stdtypes.html, Mapping Types, fetched 2026-10-11T08:50:50Z, sha256 3b6ac10c7354317fa42ed1253bf3e400c2f62af139da4a5a53ec56159d0ae763). Now the rows are kept in page order, violations() reports a need written twice and checks every row, and test_a_duplicate_source_row_is_caught_in_either_order_whatever_it_holds inserts each of the three duplicates before and after the valid row.
• One step wider, as the brief asks. page_rows() and violations() have no caller outside this module; violations() still takes a dict, so the 209-mutant test is unchanged and all 209 are still rejected. A need written twice in the mechanism table is now an assertion of the first test (len(set(needs))); a missing row and a row for no need were violations already.

Page variant (one Completion gate row changed)	f085a62b (6 tests)	be54c82ef (11 tests)
the page as it stands (control)	OK	OK (skipped=1)
wrong host replaces the valid row	FAILED	FAILED
wrong host before the valid row	OK	FAILED
wrong host after the valid row	FAILED	FAILED
trailing prose replaces the valid row	FAILED	FAILED
trailing prose before the valid row	OK	FAILED
trailing prose after the valid row	FAILED	FAILED
an exact copy before the valid row	OK	FAILED
the lines artifact without the hooks.md line 2587 entry	OK (the old guard does not read it)	FAILED

Response to the command center's Claude micro at be54c82ef (CHANGES_REQUESTED: one P2, four P3s; file research-cc-role-20261008/pr986-claude-micro-be54c82e.md, 5,523 B, sha256 52ac87180ba863c31f69233043f985ae2cfc0044524a0b2a53c0da17747e0304). The P2 and the three P3s reproduced here before any change; one forward commit, 18227acd3478656af8efb0d0023b67f76188069e, answers them. The co-op's GPT micro reads f085a62b..18227acd3 and the command center's Claude micros follow on this head.

• P2: a corrupt source row the guard never saw. page_rows() kept a source row only if the physical line started with | , so a wrong-host copy of the Completion gate row written with one leading space, three leading spaces, no space after the leading pipe, a tab after the pipe, no leading pipe or no outer pipes (each is a table row for GitHub's renderer) passed all 11 tests. At 18227acd3 the page was read the way the GFM specification reads a table (superseded by the canonical structure at 23fa8ef7a, below; https://github.github.com/gfm/, version 0.29-gfm of 2019-04-06, section 4.10, fetched 2026-10-11T09:47:16Z, sha256 b153d814fdfc8624bb6da7449162c1cd707a637f7d1c1b636eb44b9cf63fa220): "The header row must match the delimiter row in the number of cells. If not, a table will not be recognized", "The table is broken at the first empty line, or beginning of another block-level structure" (examples 201 and 202), and every other line is a row; a line is cut at a newline or a carriage return only (section 2.1). What the renderer does was measured, not assumed: gh api markdown -f mode=gfm on the header, the delimiter, the valid row and one line of each of 39 shapes before it and after it, and on 12 table starts (retained, gfm-render-measurement.json, sha256 a6e5a13dae945734eac60beba901e28f33b4659cf436a4b2f86509dc3bf3e4ba): the 12 row shapes (the canonical row, nine variants of its pipes and spaces, a third cell, and the need alone as a plain line of text) and 7 odd rows are rows; a block quote, a heading, a list item (three markers, two numbered forms), a fence (two kinds), an HTML block or comment, a thematic break (two kinds), a line indented four columns or a tab, a lone pipe and a blank line end the table, and a table that is ended before a row takes that row and every row after it out of the table. The module's reading agrees with the renderer on every shape (0 disagreements; GFM_RENDER_CHECK=1 repeats it). A line that starts with a pipe outside a table is reported, because the renderer shows it as text; a cell is trimmed of spaces and tabs only, so a no-break space in a cell is part of it; an extra cell the renderer drops (example 204) is still read. One step wider than the micro: text after a U+2028, U+0085, vertical tab, form feed or file separator inside a row was hidden from the old reading, which cut lines wherever str.splitlines() does (a no-break space after the cell was hidden by strip() the same way).
• P3-1, sed parity. For a line that is followed by a newline sed -n '<n>p' <capture> | sha256sum prints the recorded hash; for a final line without one, line_digest() adds a newline that sed does not print. The four places (the module docstring and test comment, the lines artifact's note, the page) say "a line followed by a newline", the captures test asserts that each recorded line is, and test_sed_prints_the_recorded_hash_of_a_line_followed_by_a_newline runs sed on a temporary file.
• P3-2, the locator counts. The earlier text said 256; claims.json (blob cb0fe6c39497d0f5c6261cd97315d1117acc3a86) holds 257 documentation locators, 256 with a capture hash and 253 with lines (the one without a capture hash is the skills.md locator). The module and this description say so, and test_the_locator_counts_the_module_states_are_the_counts_of_the_blob_it_names reads the blob the module names and compares the counts its docstring states.
• P3-3, an untrue sentence of this description. "The same corruptions after the valid row, or in its place, fail" was untrue for the exact copy (it passes in either position at f085a62b); the sentence above is corrected, and the table was right.
• P3-4, the dict citation. The documentation of d[key] = value does not say that an existing entry is replaced. The Python 3.13 language reference does ("This can either replace an existing key/value pair with the same key value, or insert a new key/value pair", Assignment statements, https://docs.python.org/3.13/reference/simple_stmts.html, fetched 2026-10-11T09:47:16Z, sha256 1f9cff2314892731150c210ba7e14ce362500767439d073fbb842b13c4f804b5) and so does the interpreter: CPython v3.13.16 Objects/dictobject.c (sha256 60ff1a61f3e8a6fdea5e645aa9343a87aa1aea24464159c6c965ef56c3106695), dict_ass_sub (line 3400) calls PyDict_SetItem (2578), _PyDict_SetItem_Take2 (2562) and insertdict (2558, defined at 1832), which finds the existing key (old_value, lines 1859 to 1875) and stores the new value over it with STORE_VALUE(ep, value) (lines 1885 and 1890). Both are retained beside the other pages and cited in the module's docstring.

Corrupt Completion gate row (a wrong host), placed before or after the valid row (the same result in both positions)	be54c82ef (11 tests)	18227acd3 (24 tests)
one leading space	OK	FAILED
three leading spaces	OK	FAILED
no space after the leading pipe	OK	FAILED
a tab after the leading pipe	OK	FAILED
no leading pipe	OK	FAILED
no outer pipes	OK	FAILED
the canonical spelling (control)	FAILED	FAILED
a block quote, a heading, four leading spaces or an empty line placed where the table should continue (it ends the table and takes the rows after it out)	OK	FAILED

At 18227acd3, each fix undone alone failed its own tests (retained output sha256 5246fbd8ac263d06635d9c7a584cceecf295d53ed9c379d5096e99fdb8ae34cd): the reading back to the physical lines fails 92 subtests and 2 errors, the old sed sentence in the page and the artifact 2, the docstring counts 1, the line_digest sentence 1. 36 single-edit mutants of that reading were caught (table sha256 06ec435245f0b849f54a1a053623a05479935fadb8802394b3326338d8ea4bbe; patterns, bounds, the block starts that end a table, the trimming, the line splitting, the stray-line report and the doc statements).

Response to the command center's Claude micro at 18227acd3 (CHANGES_REQUESTED: two P2s, three P3s; file research-cc-role-20261008/pr986-claude-micro-18227acd.md, 8,701 B, sha256 6152b0c4a80e4dd1d1616b61ea04b353c85d070b9ebe8d00800a3511882836e9) and to Codex's two threads of the same head (10:46:06Z and 10:46:07Z, both P2). Both were reproduced here before any change; one forward commit, 23fa8ef7a4589de6178156533c286f503dab2518, leased on 18227acd3, answers them. The command center ruled option 1: fail closed on the page's canonical structure and do not emulate a renderer. The other option it measured (markdown-it-py 4.2.0) is not adopted: it would add a CI dependency and still needs a scope rule. The co-op's GPT micro reads f085a62b..23fa8ef7a.

• P2-1 and P2-2, one root cause (Codex :90 and :223). The reading of 18227acd3 copied the GFM table rules line by line and let one pattern (OTHER_BLOCK) stand for the block structure around a table, with no container, fence or HTML-block context, so it disagreed with GitHub's renderer in both directions. Rows GitHub renders were dropped without a report: a row led by an inline tag (<b>, <span>, </b>), an autolink, a backtick fence whose info string holds a backtick, or <! (GFM 0.29-gfm, section 4.5: "If the info string comes after a backtick fence, it may not contain any backtick characters"; section 4.6, start condition 7: a complete tag "followed only by whitespace or the end of the line"). Tables that GitHub does not render as the source table were counted as it: a source row moved into a second table, a fenced block, a tilde fence, an HTML comment, a block quote (section 5.1) or a list item (section 5.2), and the whole source table inside an HTML comment, which left the published page with no source table while the module passed.
• The fix. read_page() requires the page's one shape: the heading ## Native mechanism per client, a blank line, the header, the delimiter, the rows, a blank line; then the paragraph that begins Claude column sources: (plain text), a blank line, the header, the delimiter and the rows to the end of the file. A row is | <need> | <cell> | (three cells in the mechanism table), with single spaces around each pipe and no pipe inside a cell; a table ends at the first line that is not a row; the page has newline line endings only; nothing before the mechanism heading starts a fence or raw HTML; and a delimiter-shaped or header-shaped line anywhere else in the file is a violation. Every other line of either section is reported with its line number. Nothing in the module says what GitHub does with a line.
• GitHub stays the oracle, measured. gh api markdown -f mode=gfm (gh 2.102.0) renders the canonical page as exactly two tables of 7 body rows each (headers Need | Claude Code 2.1.295 | Codex 0.161.0 and Need | Claude Code source). Of the 18 regression pages, 16 render differently (a table gone, a third table, a row hidden from a table or added to it) and 2 render the same as the canonical page (an HTML comment under the last source row; one more blank line at the end of the page); all 18 are rejected, because none is the canonical page. The renders are retained (gfm-render-regression-pages.json, sha256 25c749ce1a94429716d7a486c787eaedb57589532b09e6abc69952b324dabdfb) and test_the_canonical_page_renders_exactly_the_two_expected_tables and test_a_page_that_hides_or_adds_a_source_row_renders_differently repeat them under GFM_RENDER_CHECK=1.

Regression page	18227acd3 reports	this head reports	GitHub renders it differently
a wrong-host row led by <b> under the last source row	nothing	1	yes
a wrong-host row led by <span> under the last source row	nothing	1	yes
a wrong-host row led by </b> under the last source row	nothing	1	yes
an autolink-led row under the last source row	nothing	1	yes
a backtick-fence line with a backtick in its info string, as a row	nothing	1	yes
a line led by <! , as a row	nothing	1	yes
a <b> row for a new need under the mechanism table	nothing	1	yes
the whole source table in an HTML comment	nothing	22	yes
the Completion gate row moved to a second table	nothing	8	yes
the Completion gate row moved into a fenced block	nothing	10	yes
the Completion gate row moved into a tilde fence	nothing	10	yes
the Completion gate row moved into an HTML comment	2	10	yes
the Completion gate row moved into a block quote table	1	8	yes
the Completion gate row moved into a list item table	3	8	yes
a quoted copy of the source table with a wrong-host row	nothing	13	yes
a list-item copy of the source table with a wrong-host row	8	13	yes
an HTML comment under the last source row (invisible, still not canonical)	nothing	1	no (invisible)
one more blank line at the end of the page (invisible, still not canonical)	nothing	1	no (invisible)

The guard of 18227acd3 reports nothing on 12 of the 16 pages that GitHub renders differently (retained output sha256 ebe5a0518400dfeffc134a711cf9059f0a9cfed3630d72834e2563d115056e58, from violations() of both modules on the same page text); it reports 1 to 8 violations on the other 4. The page as written passes both.

• P3-1, this description: refreshed for this head (the scope, the counts, the response, the sources, the commands).
• P3-2, the equivalents of the surviving mutants. A lone carriage return, a copy of a row indented three spaces and a row that ends in an escaped pipe are violations (test_a_carriage_return_a_three_space_indent_and_a_row_ending_in_an_escaped_pipe_are_violations; test_a_carriage_return_in_the_prose_hides_no_table_line covers a table line that starts after a lone carriage return). 26 single-edit mutants of the canonical reading are caught (table sha256 7fe0faba6021f4383a250b05f731801841cceb843f5d5bf394e965b30f38acbb): the first run caught 15, and each of the 11 survivors (the blank-line checks, the mechanism header and row shapes, the cell pattern, the carriage return, the unreadable-structure report, the Unicode trimming) got a test.
• Observations, no change: a one-column table cannot carry a two-cell row; a footnote is flagged, which is the safe direction.

Response to the command center's Claude micro at 23fa8ef7a (CHANGES_REQUESTED: three P2s; file research-cc-role-20261008/pr986-claude-micro-23fa8ef7.md, 1,618 B, sha256 e8c4cb94fe891c852a46f87bdbdd6838bcc0b0d5e6e6c5392683b9dac55a6dc7) and to its adjudication (2026-10-11T13:50Z, the final fix round) of Codex's two P1 threads of 13:35:12Z. One forward commit, a7dac4fdf4cc0da6dff62469d3cc0bfabbc76a6f, on 23fa8ef7a; the range for the confirmation micro is 23fa8ef7a..a7dac4fdf. Every item is on lines this PR added or changed and was reproduced on 23fa8ef7a before the change.

• P2-1, a regression: a table without outer pipes inside the source paragraph. The paragraph lines skipped the sweep, so a header line, a delimiter row and a row written without outer pipes passed, and GitHub renders a third table there (GFM 0.29-gfm section 4.10: "A leading and trailing pipe is also recommended for clarity of reading, and if there's otherwise parsing ambiguity", sha256 b153d814fdfc8624…; the command center's cmark-gfm references, extensions/table.c, are its own measurement and are not re-read here). The paragraph check now also rejects a delimiter-shaped and a header-shaped line.
• P2-2: a delimiter row spaced with a vertical tab or a form feed. Both are whitespace to the renderer (GFM section 2.1: "line tabulation (U+000B), form feed (U+000C)") and neither is in the space and tab classes of the patterns, so a delimiter row spaced with one passed and GitHub rendered a table. The page rejects both characters, line by line.
• P2-3, adjudicated as fix-now: an HTML table in prose. Raw HTML is rendered as written (GFM section 6.10: "Text between < and > that looks like an HTML tag is parsed as a raw HTML tag and will be rendered in HTML without escaping"). A table tag (table, thead, tbody, tfoot, tr, th, td, any case) is rejected anywhere, in a line of prose and in a cell.
• The regression pages and the measurements. Nine pages join the regression pages (27 in all): four pipe-less tables in the paragraph (a header like the source table's, another header, alignment colons, two more lines of prose before the header), a delimiter row spaced with a vertical tab and one with a form feed, an inline HTML table in the paragraph, an uppercase HTML table inside a cell of the mechanism table and a source-header copy that renders as text. GitHub's renderer (gh api markdown -f mode=gfm, gh 2.102.0) shows 24 of the 27 differently from the canonical page and 3 identically (retained gfm-render-regression-pages-v2.json, sha256 4d130c8063339eb3685275a3de82eaa7d06086a3fde474eb799e09d5f800a427; the first file, 18 pages, is kept); all 27 are rejected. The guard of 23fa8ef7a passes 8 of the new pages with no violation (retained fail-before-wave.txt, sha256 fb293f401f13d379d08ff32aee24880346ed82cfdf2486c0506a1d744dfa4736). 32 single-edit mutants of the reading are caught, 6 of them new (table sha256 a9cb9daa3a6ff4bd044ef4b45af780efadbdd4600ebeab96599b793de433b18e).
• primary-source-reading.json:3, "Pin the billing behavior for refused advisor calls" (Codex P1). The run text said the refused calls "carry no usage and add no advisor charge". The advisor-tool page (platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool.md, retained capture sha256 81a97b998b6d8043…, 100,799 bytes) says that a failed advisor call returns an advisor_tool_result_error and that the request does not fail (lines 385-410; too_many_requests: "The advisor sub-inference was rate-limited", line 403) and that iterations of type advisor_message are billed at the advisor model's rates (lines 1256-1293); it does not say how a refused call is billed. The sentence now says that the refused calls have no advisor_message iteration in the usage record, so no advisor usage is priced for them, cites those lines and says what the page does not state.
• primary-source-reading.json:102, "Source the advisor telemetry-field claim" (Codex P1). The speed note said the API "does not report" a speed or a service tier for an advisor iteration. The page's example advisor_message iteration lists type, model and the four token counts only (lines 1273-1280), and the page names neither speed nor service_tier anywhere; the retained usage record's 29 advisor iterations carry speed, service_tier and inference_geo as null (recounted from the record, sha256 7b81e56efef8ca52…). The note says exactly that.
• The structured block and its tests. usage.advisor_page records the page (url, sha256, bytes), the cited line ranges with six quoted lines, what the page does not state, the example iteration's fields and the record's counts (29 iterations, none with a speed, a tier or a geography). test_the_advisor_claims_rest_on_the_page_lines_and_the_record_counts compares the block with the rest of the reading and builds the cited sentences from the block's line numbers, so they cannot drift; test_the_cited_page_lines_say_what_the_block_quotes (the directory with the retained page in PLATFORM_DOCS_CAPTURES) checks the hash, the size, each quote on its line, the example's fields and that the page names neither field. Both fail on the data of 23fa8ef7a (retained fail-before-advisor-claims.txt, sha256 cc634f4cf39e9529a723cecd9a7a5f0125952d94870fc5906f3fb4153a9319d6); 17 single-edit mutants of the block, the sentences and the counts are caught and the control passes (table sha256 5be91dddab3fb1c8c3b10d2cc2acef1a645dc5718fdb4901a118373de66f2ed8). The record itself stays private; the block is the committed statement of what was counted in it.
• The six invisible characters (P3-3, fix-now). Four tests of tests/test_native_practice_cross_client_sources.py carried literal no-break spaces, a zero-width space and a line separator (U+00A0 x4, U+200B x1, U+2028 x1) where they meant escapes, because an escape typed into a tool call is decoded to the character itself (measured: printf of an escape piped to xxd prints the bytes c2a0). They are escapes again and test_the_source_of_this_module_spells_its_invisible_characters_as_escapes fails on the bytes of 23fa8ef7a (['U+00A0', 'U+200B', 'U+2028'] != []; retained fail-before-escapes.txt, sha256 2249fe5826f3076b1da6c165f0a43331437e8bdbdeefb7d0e5b6b5ba81b2d242). U+2028 is a str.splitlines() boundary (Python 3.13, Built-in Types, str.splitlines). The added lines of the commit hold no character outside printable ASCII.
• The command center's confirmation micro of a7dac4fdf: PASS_WITH_P3, no correction owed (research-cc-role-20261008/pr986-claude-micro-a7dac4fd.md, 13,127 B, sha256 a32b13b5f4a85e19…). P2-1 to P2-3 are closed against the cmark-gfm mirror and GitHub's renderer, every new regression fails on the guard of 23fa8ef7a, the advisor narrowing is true at the retained page and the private record, and the range holds no P1 and no P2. Codex reviewed a7dac4fdf at 15:36:04Z ("Didn't find any major issues"). Its findings that are not blockers are in the Follow-ups below.

SOTA sources

• The GPT read of Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources #986 at 428fcb19 (CHANGES_REQUESTED, one P2), full verdict sha256 59ec9fecb2ecaae153856a83f1a2fc1767dfb3134f212e8f7f1f7d3f91349049: the finding and its measurement; reproduced here from the page and the module, not copied.
• The command center's Claude micro of Ground the Claude Code native practice in primary sources and one cross-client mapping #958 at 7b216145 (PASS_WITH_P3) and its read at 8ca5d825: the four P3s and the cross-client source request; the figures are recounted here from the retained child-usage record (sha256 7b81e56efef8ca52ae778330f6882747f6f14fff5717cb3729b37d3258386d87), not copied from either read.
• The pricing and advisor-tool pages of the advisor-spend commit, re-fetched by the script that now writes the read time (page bytes sha256 67904e280b595d31f373c857844ed380072ae79d280d8f385640b5992d337ca9 and 81a97b998b6d8043032a656214e34ffb474b3f4fa37d2724155df396ffefaf5a, unchanged).
• The Claude Code documentation pages behind each cell of the new source table: sub-agents, workflows, hooks, cli-reference, plugins, commands and headless, each at the capture claims.json already records (URL, capture sha256 and fetch time), with the cited line checked against the retained copy before writing.
• CPython 3.13.16 (the interpreter used here), Lib/re/__init__.py lines 169-172 (fullmatch returns _compile(pattern, flags).fullmatch(string); docstring "Try to apply the pattern to all of the string"; file sha256 dbe158a677c6aaacf717ea2abc23c56233453d38024aef75b7c3d93612cabb93), and the re documentation of that release, https://docs.python.org/3.13/library/re.html (re.fullmatch and the entries for . and the greedy +; fetched 2026-10-11T03:48:42Z, sha256 19195ec76056c9b8088394ba81014ff47e1d4040a8eae092e450412f75f5c378): the mechanism the guard's note group depends on.
• The GPT micro of Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources #986 at cc454f7f (CHANGES_REQUESTED, one P2, one P3), verdict sha256 a431713f07833974dd681bdb13d0f8973fe8df5a427f2004c2be029751aaa45f: the finding and its measurement, reproduced here from the page and the module, not copied.
• The Codex review in GitHub of Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources #986 at f085a62b (P2, one thread on the page's Claude column sources paragraph): the finding, reproduced from claims.json (257 documentation locators, 256 with a capture hash, 253 with lines, 2 of the 11 citations covered) before the change.
• The GPT micro of Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources #986 at f085a62b (CHANGES_REQUESTED, one P2), verdict sha256 57b8e7b253b4b1c9facd255ef883c0426f7e50df33952f0c5c19e179c90addbf: the finding and its probe pages, reproduced here (the table above), not copied.
• GitHub Flavored Markdown Spec, version 0.29-gfm (2019-04-06), section 4.10 Tables (extension) and section 2.1 (a line ending, a blank line), https://github.github.com/gfm/ (fetched 2026-10-11T09:47:16Z, sha256 b153d814fdfc8624bb6da7449162c1cd707a637f7d1c1b636eb44b9cf63fa220), and GitHub's renderer through gh api markdown -f mode=gfm (measurement file sha256 a6e5a13dae945734eac60beba901e28f33b4659cf436a4b2f86509dc3bf3e4ba): the mechanism the table reading follows.
• Python 3.13 documentation, Reference, Assignment statements (https://docs.python.org/3.13/reference/simple_stmts.html, fetched 2026-10-11T09:47:16Z, sha256 1f9cff2314892731150c210ba7e14ce362500767439d073fbb842b13c4f804b5) and CPython v3.13.16 Objects/dictobject.c lines 1832-1894 and 2538-2562 (sha256 60ff1a61f3e8a6fdea5e645aa9343a87aa1aea24464159c6c965ef56c3106695): a mapping assignment replaces an existing entry.
• The command center's Claude micro of Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources #986 at be54c82ef (CHANGES_REQUESTED: one P2, four P3s), file sha256 52ac87180ba863c31f69233043f985ae2cfc0044524a0b2a53c0da17747e0304: the findings, reproduced here from the page and the module, not copied.
• Python 3.13 documentation, Built-in Types, Mapping Types: d[key] = value, "Set d[key] to value" (https://docs.python.org/3.13/library/stdtypes.html, fetched 2026-10-11T08:50:50Z, sha256 3b6ac10c7354317fa42ed1253bf3e400c2f62af139da4a5a53ec56159d0ae763), retained with the other cited bytes.
• The command center's Claude micro of Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources #986 at 18227acd3 (CHANGES_REQUESTED: two P2s, three P3s), file sha256 6152b0c4a80e4dd1d1616b61ea04b353c85d070b9ebe8d00800a3511882836e9: the findings and the ruling (option 1), reproduced here from the page and the module, not copied. Its cmark-gfm 0.29.0.gfm.13 line references and its markdown-it-py 4.2.0 agreement count (26 of 29 samples) are the command center's measurements and are not re-read here.
• The Codex review in GitHub of Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources #986 at 18227acd3 (two P2 threads, :90 and :223): the findings, reproduced here from the page and the module before the change.
• GitHub Flavored Markdown Spec, version 0.29-gfm (2019-04-06), https://github.github.com/gfm/ (fetched 2026-10-11T09:47:16Z, sha256 b153d814fdfc8624bb6da7449162c1cd707a637f7d1c1b636eb44b9cf63fa220): section 4.5 (fenced code blocks: the info string of a backtick fence), 4.6 (HTML blocks, start condition 7), 5.1 (block quotes), 5.2 (list items) and 4.10 (tables), the mechanisms behind the 16 pages that GitHub renders differently; and GitHub's renderer through gh api markdown -f mode=gfm on those pages (gfm-render-regression-pages.json, sha256 25c749ce1a94429716d7a486c787eaedb57589532b09e6abc69952b324dabdfb).
• The command center's Claude micro of Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources #986 at 23fa8ef7a (CHANGES_REQUESTED: three P2s, follow-ups FU-1 and P3-1 to P3-5), file sha256 e8c4cb94fe891c852a46f87bdbdd6838bcc0b0d5e6e6c5392683b9dac55a6dc7, and its adjudication of 2026-10-11T13:50Z: the findings, reproduced here from the page and the module before the change. GitHub Flavored Markdown Spec 0.29-gfm (2019-04-06), https://github.github.com/gfm/ (fetched 2026-10-11T09:47:16Z, sha256 b153d814fdfc8624bb6da7449162c1cd707a637f7d1c1b636eb44b9cf63fa220): section 4.10 (the outer pipes of a table are only recommended), section 2.1 (whitespace includes U+000B and U+000C) and section 6.10 (raw HTML).
• The Codex review in GitHub of Follow-up to #958: the advisor-spend record's four P3s and the cross-client page's Claude sources #986 at 23fa8ef7a (two P1 threads, primary-source-reading.json:3 and :102, 13:35:12Z): the findings, reproduced here from the artifact, the tests and the retained page before the change.
• The advisor tool page, https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool.md (retained, 100,799 bytes, sha256 81a97b998b6d8043032a656214e34ffb474b3f4fa37d2724155df396ffefaf5a, the hash the reading already records): "Error results" (lines 385-410), "Usage and billing" (lines 1256-1293) and the example iteration (lines 1273-1280), cited by line in usage.advisor_page; and the retained usage record child-usage-wf_bee0b670-035.json (sha256 7b81e56efef8ca52ae778330f6882747f6f14fff5717cb3729b37d3258386d87, not committed), recounted for the 29 advisor iterations' fields.
• Python 3.13 documentation, Built-in Types, str.splitlines (https://docs.python.org/3.13/library/stdtypes.html, fetched 2026-10-11T08:50:50Z, sha256 3b6ac10c7354317fa42ed1253bf3e400c2f62af139da4a5a53ec56159d0ae763): the line boundaries include U+2028 and U+0085, the reason the module's source keeps such characters as escapes.

Evidence-class table

Claim	Evidence class	Command / receipt
The refused advisor calls carry no usage: the run text says only the answered calls bill	local_integration	tests.test_primary_source_reading_counts.ReadingCountsTests.test_only_the_answered_calls_are_billed_and_the_run_text_says_so; fails on #958's artifacts
The pricing read time is the clock reading of a script that re-fetched both pages and checked their hashes (unchanged), not a typed time	local_integration	test_the_pricing_read_time_is_a_clock_reading_on_the_date_the_addendum_names; fails on #958's artifacts
The list price assumes standard speed: 677 executor messages record it, 24 record none (service tier standard, $0.85 of the executors' spend at standard list) and 29 advisor iterations record neither speed nor tier	local_integration (recounted from the retained record)	test_the_list_price_states_the_speed_it_assumes; fails on #958's artifacts
The 5-minute plus 1-hour cache writes add up for the advisor's entries too, not only the executors'	local_integration	test_the_published_spend_recomputes_from_the_recorded_tokens_and_rates
Every need of the cross-client mechanism table has a Claude source cell whose citations are each a documentation page, its recorded capture hash and fetch time and a line; a changed, malformed or removed citation (also one of several), a moved line or a removed row fails	local_integration (the cells were asserted against the retained copies and claims.json when written; the test needs repository files only)	tests.test_native_practice_cross_client_sources (35 tests); every citation is validated on its own: 154 single-citation mutants fail, the first version passed 75 of them (probe table sha256 4e3a48f56a494b71f23236f58ef060712d13b5244d2255fb79cd1d8add31d8c1)
A citation added after a note ( (extra prose), the citation after a space, with no separator, after ; see , the same on another host) is rejected; a note ends at its own closing parenthesis	local_integration	tests.test_native_practice_cross_client_sources (35 tests, 209 mutants); on the old (.+) note group the new tests fail (62 failing subtests), reproduced by running a copy of the module with the old group
The guard's correction cites the Python mechanism at its pin	source_review	the module's docstring: CPython 3.13.16 Lib/re/__init__.py:169-172 and the re documentation (fetched 2026-10-11T03:48:42Z, sha256 19195ec7...), both read here
The run text and the pricing block carry the command center's two P3s	local_integration	the two assertions of tests/test_primary_source_reading_counts.py (fail on the artifact without the change)
Each cited line is recorded with the capture it was read from and the sha256 of its text, and the page, the lines artifact and claims.json agree	local_integration (the hashes were computed from the retained captures; CI checks repository text, and a host that has the captures recomputes them)	test_every_cited_line_is_recorded_with_its_capture_and_a_line_hash, test_a_line_that_is_not_recorded_or_recorded_with_another_capture_is_caught, test_the_line_hash_is_the_hash_of_the_line_and_its_newline, test_the_recorded_line_hashes_match_the_captures_where_they_are_available; the first fails on f085a62b (the artifact does not exist there)
A duplicate source row, corrupt or an exact copy, before or after the valid row, fails	local_integration	test_a_duplicate_source_row_is_caught_in_either_order_whatever_it_holds; the three pages with the duplicate first pass the module at f085a62b
The page has one canonical structure and every other line of its two sections is a violation, so a source row cannot be hidden, moved or added where GitHub renders it differently	local_integration; native_proven for the renderer (the canonical page and 18 regression pages rendered with gh api markdown, gh 2.102.0)	CanonicalStructureTests (24 tests): 12 of the 16 pages GitHub renders differently pass the violations() of 18227acd3 with nothing reported and every one of the 18 is rejected here (retained output ebe5a051…); 26 of 26 single-edit mutants of the reading caught (7fe0faba…); the live tests under GFM_RENDER_CHECK=1
A table without outer pipes in the source paragraph, a delimiter row spaced with a vertical tab or a form feed and an HTML table tag are violations: nine more regression pages (27 in all, 24 of which GitHub renders differently from the canonical page)	local_integration; native_proven for the renderer (gh api markdown, gh 2.102.0)	CanonicalStructureTests; the guard of 23fa8ef7a passes 8 of the new pages with no violation (retained fail-before-wave.txt, sha256 fb293f40...); 32 of 32 mutants caught (a9cb9daa...); the renders: gfm-render-regression-pages-v2.json, sha256 4d130c80...
The module's source holds no invisible or non-ASCII character	local_integration	test_the_source_of_this_module_spells_its_invisible_characters_as_escapes; it fails on the bytes of 23fa8ef7a (retained fail-before-escapes.txt, sha256 2249fe58...) and passes on this head
The run text and the speed note say no more than the advisor-tool page says and cite its lines: a refused call has no advisor_message iteration in the record, the page does not say how it is billed, and the record's 29 advisor iterations carry no speed or service tier	local_integration; source_review (the retained page, sha256 81a97b99...)	test_the_advisor_claims_rest_on_the_page_lines_and_the_record_counts and the gated test_the_cited_page_lines_say_what_the_block_quotes; both fail on the data of 23fa8ef7a (retained fail-before-advisor-claims.txt, sha256 cc634f4c...); 17 of 17 mutants caught (5be91ddd...)
The statements about the artifacts are the artifacts' counts: 257 locators, 256 with a capture hash, 253 with lines; the sed parity is for a line followed by a newline	local_integration	test_the_locator_counts_the_module_states_are_the_counts_of_the_blob_it_names, test_sed_prints_the_recorded_hash_of_a_line_followed_by_a_newline and the captures test (every recorded line is newline-terminated); each fails when its fix is undone alone

Declared test contract change

The eighth commit (a7dac4fdf) changes two existing tests of tests/test_primary_source_reading_counts.py to the narrowed sentences and makes them stronger, not weaker: test_only_the_answered_calls_are_billed_and_the_run_text_says_so now requires the sentence about the advisor_tool_result_error and the advisor_message iteration and rejects "add no advisor charge", and test_the_list_price_states_the_speed_it_assumes now requires "carry neither speed nor service tier" and rejects "which the API does not report". Two tests are added there (the block against the rest of the reading, and the gated check against the retained page). In tests/test_native_practice_cross_client_sources.py (33 tests become 35) two tests are added (the module's source is printable ASCII; the findings for a table without outer pipes, a control space and an HTML table tag), nine regression pages join regression_pages(), and no assertion is dropped.

The third forward commit (23fa8ef7a) replaces that reading with the page's canonical structure (the command center's ruling, option 1) and declares this change to the module: SourceTableReadingTests (13 tests) and unfixed_rows() are gone, CanonicalStructureTests (22 tests) takes their place, and twenty-four tests become thirty-three. Four of the 13 survive under their names (the Unicode line-separator case, the non-space characters in a cell and the sed parity are rewritten to read through the new reading; the locator-count test is unchanged). Nine are removed, and each assertion they made is either carried or reversed on purpose: test_what_starts_a_table and test_the_reading_agrees_with_the_rows_github_renders asserted the emulation of GFM table starts and 39 line shapes, and are replaced by the two renderer tests that render the canonical page and the regression pages; test_a_corrupt_source_row_in_any_spelling_github_renders_as_a_row_is_caught_in_either_order, test_a_line_that_ends_the_table_takes_the_rows_after_it_out_of_the_table, test_a_second_table_a_stray_row_and_a_table_that_lost_its_delimiter_are_caught, test_the_page_as_written_is_one_table_of_seven_rows_and_one_of_seven_needs and test_a_blank_line_is_only_spaces_and_tabs are carried by test_every_line_the_canonical_page_does_not_have_is_a_violation_wherever_it_is_put, test_the_row_shapes_are_exact, test_a_table_ends_at_the_first_line_that_is_not_a_row_in_both_tables, test_a_source_row_moved_out_of_the_table_is_reported_wherever_it_goes, test_a_header_or_delimiter_that_is_not_the_canonical_one_is_a_violation, test_the_findings_name_what_is_missing_from_the_structure and test_the_page_as_written_has_the_canonical_structure; test_an_extra_cell_the_renderer_drops_is_still_read and test_lines_that_only_look_like_something_else_stay_rows asserted that a non-canonical shape is read as a row, and that is reversed: such a line is a violation now. Every regression page and the 209 citation mutants are rejected.

The second forward commit (18227acd3) made the reading of the two tables stricter and dropped no assertion: page_rows() read each table by structure (header, delimiter, rows until the table ends) instead of by the leading | , so a line that starts with a pipe outside a table became a violation, and the old reading was kept in the module as unfixed_rows() only to show what it missed (it is gone at 23fa8ef7a). split_cell() trims spaces and tabs only. Eleven tests became twenty-four.

The first change: One test of the new module is replaced by stronger ones: test_a_changed_citation_is_caught (it changed both Completion gate citations together) becomes test_every_citation_of_every_cell_is_caught_when_changed_or_removed_by_itself (154 mutants, each citation by itself; 209 after the micro, with the five probes of the second read as a test of their own) and test_the_probes_of_the_review_of_428fcb19_are_caught, with test_a_missing_cell_a_new_need_and_a_reordered_cell_are_caught added; nothing it asserted is dropped. Elsewhere tests are added and one existing test gains an assertion (the advisor's cache split); no other existing assertion changes.

Local commands run

Head f085a62b20e6add05dba0e56b497d0d0437ebffb on 03a986ed6 (CPython 3.13.16 locally), TMPDIR outside /tmp and outside home:
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11423, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11423, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; no change to a generated report
$ python3 -E -s -B -m unittest tests.test_primary_source_reading_counts tests.test_native_practice_snapshot_provenance tests.test_native_practice_doc_corrections tests.test_native_practice_cross_client_sources tests.test_evidence_manifest tests.test_upstream_surface_watch tests.test_practice_references tests.test_record_verdicts tests.test_skills_manifest tests.test_skills_status tests.test_runtime_worker_skills tests.test_landscape_sweep_skills tests.test_install_skills
Ran 621 tests in 60.891s; OK (skipped=8)
$ a copy of tests/test_native_practice_cross_client_sources.py with the old note group `(.+)`, run as a module
FAILED (failures=62): 56 in the mutation test, 6 in the probe test
$ git diff origin/main --unified=0 | grep '^+'   # added lines, scanned for the host's private names
hits 0

Head be54c82ef067cada6124c709ac8fe674fc1bb1aa (the working tree as tested, committed unchanged), same environment:
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11424, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11424, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; no change to a generated report
$ python3 -E -s -B -m unittest tests.test_primary_source_reading_counts tests.test_native_practice_snapshot_provenance tests.test_native_practice_doc_corrections tests.test_native_practice_cross_client_sources tests.test_evidence_manifest tests.test_upstream_surface_watch tests.test_practice_references tests.test_record_verdicts tests.test_skills_manifest tests.test_skills_status tests.test_runtime_worker_skills tests.test_landscape_sweep_skills tests.test_install_skills
Ran 626 tests in 54.309s; OK (skipped=9)
$ CLAUDE_DOCS_CAPTURES=<the retained documentation pages> python3 -E -s -B -m unittest tests.test_native_practice_cross_client_sources
Ran 11 tests; OK (the capture test runs: every recorded line hash is recomputed from the bytes)
$ the old module and the new one against eight pages that change the Completion gate source row (the table above)
the old module passes the three duplicates-first pages and ignores the lines artifact; the new module fails every page but the control
$ git diff HEAD~1 --unified=0 | grep '^+'; cat cross-client-lines.json   # added lines, scanned for the host's private names
hits 0

Head 18227acd3478656af8efb0d0023b67f76188069e (the working tree as tested, committed unchanged), same environment, run after `scripts/host_receipts.register_file` had registered the three changed files:
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11424, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11424, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; the only changes are the 4 files of the commit
$ python3 -E -s -B -m unittest <the same 13 modules>
Ran 639 tests in 53.599s; OK (skipped=9)   (rc 0)
$ CLAUDE_DOCS_CAPTURES=<the retained documentation pages> GFM_RENDER_CHECK=1 python3 -E -s -B -m unittest tests.test_native_practice_cross_client_sources
Ran 24 tests; OK (both gated tests run: every recorded line hash is recomputed from the bytes, and the module's reading agrees with GitHub's renderer on 39 line shapes and 12 table starts)
$ the module of be54c82ef and this head against 22 corrupted pages (the table above), and each fix undone alone (retained output 5246fbd8...)
the old module passes the 20 pages that are not the control; this head fails all 22
$ 36 single-edit mutants of the table reading
36 of 36 caught (table sha256 06ec4352...)
$ git diff HEAD~1 --unified=0 | grep '^+'   # added lines, scanned for the host's private names
hits 0

Head 23fa8ef7a4589de6178156533c286f503dab2518 (the working tree as tested, committed unchanged), same environment, run after `scripts/host_receipts.register_file` had registered the two changed files:
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11424, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11424, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; no change to a generated report
$ python3 -E -s -B -m unittest <the same 13 modules>
Ran 648 tests in 51.134s; OK (skipped=10)   (rc 0)
$ CLAUDE_DOCS_CAPTURES=<the retained documentation pages> GFM_RENDER_CHECK=1 python3 -E -s -B -m unittest tests.test_native_practice_cross_client_sources
Ran 33 tests; OK (all three gated tests run: every recorded line hash is recomputed from the bytes, the canonical page renders as exactly the two expected tables, and each dangerous regression page renders differently)
$ the module of 18227acd3 and this head against the 18 regression pages (the table above)
the old module reports nothing on 12 of the 16 pages GitHub renders differently; this head rejects all 18 and passes the page as written
$ 26 single-edit mutants of the canonical reading
26 of 26 caught (table sha256 7fe0faba...)
$ git diff HEAD~1 --unified=0 | grep '^+'   # added lines, scanned for the host's private names
hits 0

Head a7dac4fdf4cc0da6dff62469d3cc0bfabbc76a6f (the working tree as tested, committed unchanged) on `23fa8ef7a`, same environment, run after `scripts/host_receipts.register_file` had registered the changed files:
$ python3 -E -s -B scripts/validate.py
{"components": 70, "hashed_files": 11424, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/evidence_manifest.py --check
{"files": 11424, "status": "passed"}   (rc 0)
$ python3 -E -s -B scripts/validate_convergence.py --all-recorded
rc 0
$ python3 -E -s -B scripts/component_matrix.py --write && python3 -E -s -B scripts/new_host_grand_list.py --write
rc 0; no change to a generated report
$ python3 -E -s -B -m unittest <the same 13 modules>
Ran 652 tests in 64.146s; OK (skipped=11)   (rc 0)
$ PLATFORM_DOCS_CAPTURES=<the retained platform pages> CLAUDE_DOCS_CAPTURES=<the retained documentation pages> GFM_RENDER_CHECK=1 python3 -E -s -B -m unittest tests.test_primary_source_reading_counts tests.test_native_practice_cross_client_sources
Ran 57 tests; OK (every gated test runs: the page lines, the recorded line hashes, and the renders of the canonical page and the 27 regression pages)
$ the module of 23fa8ef7a and this head against the 27 regression pages (fail-before-wave.txt)
the old module passes 8 dangerous pages with no violation; this head rejects all 27 and passes the page as written
$ 32 single-edit mutants of the canonical reading; 17 single-edit mutants of the advisor block, sentences and counts
32 of 32 caught (table sha256 a9cb9daa...); 17 of 17 caught, the control passes (table sha256 5be91ddd...)
$ git diff HEAD~1 --unified=0 | grep '^+'   # added lines, scanned for the host's private names and for characters outside printable ASCII
hits 0; characters outside printable ASCII: none

6 files changed, 1150 insertions(+), 9 deletions(-) at the new head (against the base). The local full-suite shards in a CI-like environment are noisy on this host (four of eight red at a CI-green head on tests no change touches), so the list above is the targeted run and CI's eight shards are the full-suite measurement.

Follow-ups

Listed under the command center's convergence ruling of 2026-10-11T13:37Z (research-cc-role-20261008/ruling-pr-wave-convergence-20261011.md) and its micro at 23fa8ef7a; none is a blocker.

• FU-1: a source note can carry another host's link; the lines come from f085a62b, so it is pre-existing.
• P3-1: the module reads the page with read_text, which normalizes CRLF, so the carriage-return finding cannot fire on the file as read from disk (read_bytes would).
• P3-2: a circular oracle (the micro's wording; the render check is gated and is the oracle of the reading's own shapes).
• P3-4 and P3-5: listed without text in the micro (pr986-claude-micro-23fa8ef7.md).
• The confirmation micro's P3 (a) 1: HTML_TABLE allows white space after < (<\s*/?\s*), so x < TD > y is rejected with the message "an HTML table tag, which the renderer shows as a table" although the renderer shows it as text (CommonMark needs the tag name right after <); the rejection fails closed and only the message and the case's label are wrong.
• The micro's P3 (a) 2: the module docstring says the guard of 23fa8ef7a passes "8 of the new pages" with no violation; it passes all 9 (8 dangerous and the invisible source-header copy). The retained fail-before-wave.txt and this description say it precisely.
• The micro's (b) 1, wider and unchanged in the range: a one-column table with no pipe, placed outside the two sections, passes (the final sweep requires a pipe in the line; the scanner makes every pipe optional), and GitHub renders a third table. A one-column table cannot carry a need/source pair and the gated canonical-render test fails on it; inside the source paragraph the same lines are rejected. And (b) 2: the oracle reader (rendered_tables) mis-attributes rows of a nested table (it resets the row at the inner <tr>), which belongs with the circular-oracle follow-up; the "renders differently" verdict still holds.
• Research queue: run the GFM render oracle (rendered_tables through gh api markdown) in CI.
• Not adopted, by the command center's ruling at 18227acd3: markdown-it-py 4.2.0 as the table parser (option 2 of its micro: it would add a CI dependency and still needs a scope rule). It stays a measured alternative.
• The advisor claims rest on the private usage record (child-usage-wf_bee0b670-035.json, not committed); usage.advisor_page is the committed statement of what was counted in it, and the gated test checks the page, not the record. A minimized extract of the 29 advisor iterations (no content) in the repository would let a test read the counts from rows; not done here.
• The command center's cmark-gfm 0.29.0.gfm.13 and 587a12bb line references are its own measurements and are not re-read here.

Decision record

No new record: the two items are corrections to #958's evidence and a source table for its cross-client page.

Host evidence

Not applicable: no file under evidence/hosts/ changes.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no floating tags). (none changed)
• New/changed workflows declare top-level permissions: {} and grant each job only what it needs. (none changed)
• No secrets are printed, logged or committed; no new required secret was added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved (not deleted, moved or overwritten).

🤖 Generated with Claude Code

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-11T15:36:05.751908Z	a7dac4f	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f085a62b20

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Another round soon, please!

Reviewed commit: be54c82ef0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 18227acd34

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 23fa8ef7a4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Hooray!

Reviewed commit: a7dac4fdf4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".