Skip to content

docs(rfc-0001): patch §§6.4, 6.7, 6.9, 9 with event-storming deltas - #9

Merged
jensholdgaard merged 2 commits into
mainfrom
docs/rfc-0001-event-storming-patches
May 10, 2026
Merged

docs(rfc-0001): patch §§6.4, 6.7, 6.9, 9 with event-storming deltas#9
jensholdgaard merged 2 commits into
mainfrom
docs/rfc-0001-event-storming-patches

Conversation

@jensholdgaard

Copy link
Copy Markdown
Owner

Summary

Four prose-only edits to RFC 0001, surfaced by an event-storming
session run on 2026-05-10. No §5 scenarios change, no test stubs
flip; status stays at red. The full session artefact lives at
scratch/event-storming-rfc-0001.md (gitignored).

What each edit does

  • §6.4 — WAL framing of audit events. RFC0001.7 says one
    attach can emit two ordered audit events (template_widened
    then template_type_expanded). New paragraph names the contract
    owed by the future ourios-wal RFC: both audit events plus the
    data record must land as one WAL framing unit, so crash recovery
    cannot observe a version bump in memory that the audit stream
    doesn't substantiate.

  • §6.7 — Alias index lifecycle. The original prose said
    "template_widened events feed the alias index." That's wrong —
    widening is intra-leaf (one template_id, several versions),
    the alias index is inter-leaf (different template_ids, semantic
    equivalence). Replaced with an explicit "alias index has no
    creation event in this RFC" paragraph naming the three candidate
    mechanisms (operator-driven, automatic-inference, deferred) and
    pointing to §9.

  • §6.9 — Replay mode. WAL replay re-walks the live attach()
    code path, which would re-fire every counter / histogram /
    gauge update. Polluting merges_total, confidence, etc. for
    the post-restart horizon. Names "replay mode" (events processed
    and state mutated as live, but §6.8 metrics suppressed) and adds
    a single replay-only wal_replay_progress gauge — documented
    inline, explicitly not added to §6.8's table so §3.1.2's
    mandatory-metric-set scenario stays unchanged.

  • §9 — Two new subsections:

    • Multi-tenancy and operational lifecycle — tenant lifecycle
      bookends (TenantInitialised, TenantPaused, TenantDeleted),
      per-tenant fairness / back-pressure pointer, alias-index
      creation mechanism (gates RFC 0002's
      template_id.resolves_to(X)).
    • Cross-RFC contracts pending — querier ↔ live template
      registry seam (RFC 0002 needs the answer before its DSL can
      compile); audit-event Parquet schema owed by a future
      ourios-parquet RFC.

Invariant / hazard impact (CLAUDE.md §§3, 4)

  • §3.1 (no silent merges). The §6.4 WAL-framing addition
    strengthens the invariant: previously the audit guarantee was
    in-memory only; now the durable contract is named.
  • §3.5 (Parquet schema migration). Unaffected — the new §9
    bullet on audit-event Parquet schema names a future contract,
    doesn't change the existing one.
  • §3.7 (multi-tenancy). §9's tenant lifecycle bullet
    acknowledges the gap §3.7 leaves around tenant deletion /
    pausing / eviction; the RFC is not committing to a mechanism,
    it is naming what it leaves to a future RFC.
  • H1 (template miner correctness). §6.4 WAL-framing tightens
    the H1 audit-event guarantee.
  • H5 (template schema evolution). §6.7 alias-index paragraph
    separates the cross-version axis from the cross-alias axis,
    removing a misleading "events feed the alias index" claim.

Test plan

  • mdbook build clean
  • cargo fmt --all --check clean
  • cargo clippy --all-targets --all-features -- -D warnings clean
  • cargo test --all-features — outer loop unchanged (3 passed,
    28 ignored; no test stubs touched by this PR)
  • No §5 scenario IDs added or removed; greppability per
    docs/verification.md §2.3 preserved

🤖 Generated with Claude Code

Surfaced by a 2026-05-10 event-storming session against RFC 0001
(scratch/event-storming-rfc-0001.md). Four prose-only edits; no §5
scenarios change, status stays at red.

§6.4 — WAL framing of audit events. RFC0001.7 says one attach can
emit two audit events in order (template_widened then
template_type_expanded); the prose now names the contract owed by
the future ourios-wal RFC: both events plus the data record must
be one WAL framing unit so crash recovery cannot observe one event
without the other. Without this, replay would bump
template_version fewer times than the in-memory leaf did, leaving
recovered state the audit stream cannot substantiate.

§6.7 — Alias index lifecycle. The original prose said
"template_widened events feed the alias index," which is wrong:
widening is intra-leaf (one template_id, several versions) and
the alias index is inter-leaf (different template_ids, semantic
equivalence). Replaced with an explicit "alias index has no
creation event in this RFC" paragraph that names the three
candidates (operator-driven, automatic-inference, deferred) and
points at §9. RFC 0002's template_id.resolves_to(X) is now
accurately described as operating against an out-of-band index.

§6.9 — Replay mode. WAL replay re-walks attach()/widen()/
expand_slot_types() against the same code path live ingest uses,
which would re-fire every counter increment, every histogram
observation, and every gauge update — polluting steady-state
metrics for the post-restart horizon (a 10-minute replay can
shift merges_total by orders of magnitude in seconds). Names
"replay mode": domain events processed and tree state mutated
exactly as live, but §6.8's metric set is suppressed. Adds a
single replay-only wal_replay_progress gauge documented inline,
explicitly NOT added to §6.8's table (so §3.1.2 mandatory-metric-
set scenario stays unchanged and the §3.1 invariant set continues
to mean steady-state observability).

§9 — Two new subsections:

  • Multi-tenancy and operational lifecycle: tenant lifecycle
    bookends (TenantInitialised, TenantPaused, TenantDeleted —
    none specified yet); per-tenant fairness / back-pressure
    pointer to a future ourios-ingester RFC; alias index
    creation mechanism gating RFC 0002.

  • Cross-RFC contracts pending: the querier ↔ live template
    registry seam (snapshot read vs RPC vs Parquet side-stream —
    RFC 0002 needs the answer before its DSL can compile);
    audit-event Parquet schema (Rust struct in §6.4 specified;
    on-disk column layout owed by a future ourios-parquet RFC).

Verification (CLAUDE.md §6.6): mdbook build clean, cargo fmt
clean, cargo clippy clean, cargo test passing (no test changes).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Updates RFC 0001 (template miner) with event-storming-derived clarifications around WAL durability contracts, alias-index semantics, WAL replay telemetry behavior, and newly surfaced open questions—without changing §5 scenarios or gate status.

Changes:

  • Add an explicit WAL atomicity/framing contract for multi-audit-event attaches (§6.4).
  • Correct/clarify alias index semantics and lifecycle (distinct from widening) (§6.7) and expand open questions (§9).
  • Define “replay mode” to prevent WAL replay from polluting steady-state metrics, with a replay-only progress gauge (§6.9).

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment thread docs/rfcs/0001-template-miner.md Outdated
Comment thread docs/rfcs/0001-template-miner.md Outdated
Comment thread docs/rfcs/0001-template-miner.md Outdated
Three Copilot hits on the event-storming patches PR, each addressed:

- C1 (§6.4 line 844): the WAL framing paragraph was internally
  inconsistent — the opening said "single framing unit" (full
  atomicity) while the fallback said "hold the data record back
  until both audit events are durable" (ordering only). Rewrite to
  state the *minimal sufficient* property: an
  "ordering-plus-durability-barrier" — a data record at
  template_version V must not become durable before every audit
  event justifying V is durable. Crash recovery may observe some
  prefix of [event_1, event_2, …, data_record], never a state in
  which the data record exists without the events that caused its
  version stamp. The framing strategy (composite frame, batched-
  fsync ordering, two-phase write-then-link) is ourios-wal's
  choice; the ordering barrier is this RFC's requirement. Removes
  the contradictory "atomicity ... or hold the data record back"
  prose.

- C2 (§6.7 line 1002): the alias paragraph named a "RecordAlias"
  command and pointed at §9 for its shape — but §9's bullet
  doesn't define RecordAlias or its shape. Speculating on an API
  name when the whole mechanism is undefined was the wrong move.
  Drop the named command; say only "the shape of the write API
  they would expose are open questions in §9." The §9 alias bullet
  is unchanged (it never mentioned RecordAlias).

- C3 (§6.9 line 1089): "the counters / histograms / gauges in
  §6.8 are suppressed" was ambiguous between "not updated" and
  "not registered/exposed." Since §3.1.2 requires the §6.8 metrics
  to be present in the registry, replay mode keeping them present
  but not updating them is the load-bearing distinction — make it
  explicit. New text: counters do not increment, histograms do not
  observe, gauges retain their previous value (or zero/empty
  initialisation); the metrics remain registered and visible to
  scrapes, only the update path is gated, not the registration
  path. §3.1.2's invariant explicitly named as preserved across
  replay.

No §5 scenarios added or removed; status stays at red. The
underlying contracts and §9 deferrals are unchanged — these are
prose precision fixes, not design changes.

Verification (CLAUDE.md §6.6): mdbook build clean, cargo fmt
clean, cargo test passing (no test changes).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@jensholdgaard
jensholdgaard merged commit 42be82a into main May 10, 2026
7 checks passed
@jensholdgaard
jensholdgaard deleted the docs/rfc-0001-event-storming-patches branch May 10, 2026 12:46
jensholdgaard added a commit that referenced this pull request Jun 6, 2026
…ed → specified (#127)

* docs(rfc-0003): specify §5 G/W/T scenarios + OTel-spec enrichments — drafted → specified

Fills the §5 acceptance-criteria stub on RFC 0003 (OTLP receiver) with
15 Given/When/Then scenarios — the 11 originally sketched plus four
enrichments surfaced by the OpenTelemetry transport-spec citations the
maintainer pulled in (empty request → success, identity+gzip MUST,
default /v1/logs path + override, concurrent Export calls).

§5.2 (Crash-before-ack) lands as **at-least-once with retry
tolerance**: the OTLP spec's *duplicate-data* section explicitly
accepts client-retry duplicates as the right tradeoff for telemetry
data, and the Collector's own WAL guidance carries the same caveat.
This resolves §9 open question #1 (dedup) — the receiver implements no
de-duplication in this RFC; any future dedup mechanism is additive.
§5.13 (compression) resolves §9 open question #9 — identity + gzip are
both required acceptance criteria; zstd/br stay out of scope (HTTP 415).

§8 is filled to map each scenario id to its test technique (proptest
for wire-decode equivalence, child-process SIGKILL harness mirroring
PR #126 for crash-before-ack, table-driven for the transport-error and
edge-OTLP arms, criterion for the latency/throughput benches).
Frontmatter status flips drafted → specified per the maturity ladder
(docs/rfcs/README.md §Lifecycle); §10 references unchanged
(the OTLP spec link already covers the cited sub-sections).

No code touched — per the split-doc-from-code-PRs convention, the §5
test stubs (red gate per docs/verification.md §3) ride a follow-up
implementation PR.

* docs(rfc-0003): address Copilot review feedback on §5 / §6.5

- §6.5: replace "dedup mechanism not yet specified" with the §9 #1
  resolution (at-least-once with duplicates per OTLP spec; any future
  dedup is additive). Removes the §6.5↔§9 internal inconsistency.
- RFC0003.2: broaden the kill-point description from "the §6.5 step
  5/6 gap" to "anywhere in the step-4-through-6 window," since the
  duplicate-on-retry contract holds across both the 4/5 and 5/6 gaps
  once the records are durable.
- RFC0003.4 / RFC0003.12: drop `wal_syncs_total` / `wal_unflushed_bytes`
  metric-name references from acceptance criteria; reframe in terms of
  observable WAL state (frame count, segment offsets, append/sync call
  counts via a test wrapper). Acceptance criteria should not pin
  telemetry identifiers.
- §5: reference `docs/rfcs/README.md` *Required sections* by heading
  name, not the malformed `§Required-sections` token.

The scenario-id grammar question (RFC0003.<m> vs §3.<n>.<m> for
invariant-tagged scenarios) is left for the maintainer — see the PR
reply on that thread.

* docs(rfc-0003): align WAL framing with RFC 0008 (per-batch, not per-record)

Copilot's second review pass caught a real contract bug: RFC 0003's §5
scenarios and §6.5 step sequence asserted per-record WAL appends, but
RFC 0008 §4 / §7.3 specifies `FrameKind::OtlpBatch` — one frame per
`ExportLogsServiceRequest`, carrying the verbatim protobuf bytes. RFC
0008 §7.3 explicitly rejects per-record framing as the alternative.

Fixes:

- §5.1, §5.2, §5.15: assertions now reference the `OtlpBatch` frame and
  its payload bytes round-tripping to the input request, not "every
  record in the WAL." The retry-duplicate assertion is now "a second
  `OtlpBatch` frame whose payload bytes equal the first."
- §6.1 step 5, §6.5 preamble, §6.5 step 3: durability language switched
  from "every record durably written" / "appends every record" to
  "batch's `OtlpBatch` frame is durably written" / "appends the encoded
  request as a single `OtlpBatch` frame."
- §8 testing-strategy crash-before-ack: asserts the same `OtlpBatch`
  payload bytes appear twice (two frames, one per export attempt).
- §5.11: drop the `max_request_bytes` config-key reference; phrase as
  "configured request-size limit" since the knob isn't specified in this
  RFC and pinning a name in acceptance criteria is premature.

The miner still consumes records per-record (§6.1 step 4–5, §9 open
questions); only the WAL frame is per-batch.

* docs(rfc-0003): close §6.5 ack-conditions and greppability gaps

Copilot's third review pass:

- §6.5: qualify the "ack only after `OtlpBatch` is durably written"
  contract as applying to non-empty batches; explicitly point at
  RFC0003.12 as the empty-batch exception (which returns success with
  no WAL write).
- RFC0003.1, RFC0003.15: lock in the §6.5 step-5 miner-acceptance
  condition alongside `Wal::sync` return. Both scenarios now assert
  ack-after-(sync AND miner-accept), per-call for RFC0003.15. RFC0003.1
  uses an instrumented `MinerCluster` stub; RFC0003.15's per-call probe
  records both orderings.
- §5 intro, §8 testing-strategy: clarify the greppability convention —
  the verbatim id reference lives in each test's leading doc comment
  (`/// Scenario RFC0003.1 — ...`) per `docs/verification.md` §2.3, not
  in the Rust function name. The `#[test] fn rfc0003_1_*` example was
  misleading.

* docs(rfc 0003): fix RFC 0008 cross-ref §4 -> §3.2 + §6.2.3

§4 of RFC 0008 is "Background — existing Rust durability ecosystem";
the FrameKind::OtlpBatch payload contract lives in §3.2 ("What goes
into the WAL") with the frame/payload-encoding details in §6.2.2 /
§6.2.3.

* docs(rfc 0003): tighten singular-frame, §6.3 error wording, decouple miner internals

Three Copilot nits:

1. RFC0003.1 said `Wal::sync` "covering the batch's frames" — singular
   per RFC 0008 §3.2 (one `OtlpBatch` frame per export batch).
2. §6.3 said the tenant-resolution error names "the failing Resource";
   RFC0003.4 names "the failing `ResourceLogs` index". Aligned both on
   the index + attribute key.
3. RFC0003.8's last `And` reached into miner internals (RFC 0001 §6.2
   step-0 short-circuit). The receiver's contract is the byte-for-byte
   pass-through; how the miner subsequently routes it is the miner's
   business. Asserted via an instrumented `MinerCluster` stub instead.

* docs(rfc 0003): broaden RFC0003.12 to all three "empty" shapes

"Empty" in OTLP can mean any of: empty resource_logs, empty
scope_logs per ResourceLogs, or empty log_records per ScopeLogs.
RFC0003.12 only covered the first shape; the fast-path contract
("zero records => success without WAL write") applies to all three.
All three are now explicit in the Given and tested.

* docs(rfc 0003): broaden unknown-fields rule, own the wire-0→None rule

1. RFC0003.6's unknown-fields clause restricted forward-compatibility
   to *top-level* fields; the OTLP/proto3 rule is "anywhere in the
   message". Reworded to "anywhere in the request body (top-level,
   nested, repeated)".
2. RFC0003.9 asserted observed_time_unix_nano=0 -> None but the rule
   was nowhere written down — RFC 0001 §6.1 just types the field as
   Option<u64> without defining the conversion. Made the contract
   explicit: the receiver owns the wire-0→None mapping, severity_number
   stays 0 because UNSPECIFIED is a valid OTLP value (not absence).

* docs(rfc 0003): respect WAL single-writer contract in replay assertions

RFC0003.1 / .4 / .15 each verified WAL state by opening a "fresh
Wal::open" while the receiver was still alive. The Wal contract
is "opened by exactly one writer" (ourios-wal/src/lib.rs §6.2 —
O_APPEND single-writer handle), so a second open while the
receiver's handle is live violates the contract and would race
with replay's heal/truncate path.

Spelled out in all three scenarios: shut the receiver down (which
drops its Wal handle) before opening a *second* Wal and replaying.

* docs(rfc 0003): use OtlpBatch frame terminology in WAL absence assertions

Three spots described "no record appended to the WAL" / "no partial
record appended" — but the WAL persistence unit is per-export
FrameKind::OtlpBatch frames, not per-record entries. Reworded
RFC0003.4, RFC0003.11, and §8's transport-errors testing bullet
to assert "no OtlpBatch frame appended" so the spec is unambiguous
about what the WAL stores.

* docs(rfc 0003): retry frames must be semantically equal, not byte-equal

RFC0003.2 and §8's crash-before-ack bullet both required the retry's
OtlpBatch frame to be byte-identical to the first. That over-constrains
clients: a retry can legitimately re-encode (JSON ordering, whitespace,
switched encoding/compression) and remain semantically the same export.

Both now assert that the second frame's payload decodes (via prost)
to an ExportLogsServiceRequest semantically equivalent to the first.

Also retitled RFC0003.13 to "Compression over HTTP: identity and gzip
MUST be supported" — the scenario is HTTP-only (Content-Encoding),
matching §9's wording.

* docs(rfc 0003): tidy §8 — Body-fork grammar + drop PR-M2 jargon

1. §8's Body-fork bullet had awkward grammar ("variants assert
   `Body::from_any_value` routes…"). Reworded as "each asserting
   that …" so the subject of the assertion is unambiguous.
2. Dropped the "PR-M2" label from the two-loop red-gate note; the
   label is internal milestone shorthand that's not defined in this
   RFC. Now refers to the two-loop pattern itself (and to RFC 0008
   §5 as the precedent) so the note is self-contained.

* docs(rfc 0003): tighten payload-equality wording in RFC0003.1 + .2

RFC0003.1's last `And` claimed "payload bytes equal the encoded
`ExportLogsServiceRequest`" but the named verification only checks a
`prost` round-trip. Protobuf has multiple wire encodings for the same
message, so the claim was strictly stronger than what the test asserts.
Relax to "payload decodes (via `prost`) to the input request" and
cross-link the RFC0003.2 explanation of why byte equality is the wrong
shape of contract here.

RFC0003.2's first `Then` said "decodes byte-for-byte" — internally
inconsistent (decoding yields a message value, not bytes). Rephrase to
"decodes (via `prost`) to an `ExportLogsServiceRequest` semantically
equivalent to the killed process's input," matching the second `And`
which already had the right shape after round 10.

Both Copilot comments (PR #127, comments 3367201831, 3367201836).

* docs(rfc 0003): disentangle ordering vs content check in RFC0003.1

Copilot caught that the last `And` clause bundled two distinct
assertions — "frame exists with the right payload" and "before the ack
fires" — but only the former is verified by the named technique
(post-shutdown replay via `Wal::replay`). Replay establishes
existence + content post-response; it can't establish pre-ack timing,
because by the time the second `Wal` is opened the response has
already shipped.

The before-the-ack ordering is what the `AtomicBool` probe in the
preceding `Then` + `And` clauses is for (set after `Wal::sync`
returns; asserted `true` by the response-writer; asserted `false` at
every pre-sync stage). That's the timing contract. The replay is the
durability/content contract. Rewording makes the split explicit so a
reader doesn't infer the replay proves the ordering.

PR #127, comment 3367208364.

* docs(rfc 0003): align §6.5 step 3 with RFC0003.1/.2 wire-encoding stance

Copilot caught that §6.5 step 3 said the WAL payload is stored as
"verbatim `ExportLogsServiceRequest` protobuf bytes" — strictly
stronger than what RFC0003.1's acceptance criteria require (rounds 12+13
relaxed those to "decodes via `prost` to the input request"). The
mismatch is most obvious on the HTTP/JSON path: there are no incoming
protobuf bytes to be "verbatim" — the receiver must encode the decoded
message to protobuf for WAL storage.

Reword step 3 to require the payload be a protobuf-encoded
`ExportLogsServiceRequest` decodable via `prost`, semantically
equivalent to the input. The wire-bytes-verbatim option remains
explicitly permitted (it is the obvious zero-copy choice for the
protobuf paths), but it is no longer the *contract*.

PR #127, comment 3367215593.

* docs(rfc 0003): cite RFC 0008 §3.1 for single-writer (not source file §6.2)

Copilot caught (three identical comments on lines 260, 340, 534) that
`crates/ourios-wal/src/lib.rs` §6.2 is a confusing cross-reference:
the file is Rust source, not a sectioned doc. The §6.2 the source-file
comment at line 162 itself cites is RFC 0008 §6.2 (segment layout),
which is where the append-only property lives — but the actual
*single-writer architecture* statement RFC 0003 is leaning on is
RFC 0008 §3.1 ("single-writer single-node component").

Rephrase all three sites to:

  per RFC 0008 §3.1's single-writer architecture
  (enforced by `crates/ourios-wal/src/lib.rs:162`)

— RFC-first as the contract; source file kept as the concrete
enforcement point. Source file gets a line-number cite, which is the
right shape for a Rust file.

PR #127, comments 3367230544, 3367230558, 3367230561.
jensholdgaard added a commit that referenced this pull request Jul 16, 2026
…gnostics

Run #9 (600s) measured 96.5% (11123/11523), statistically the same as
run #8's 97.1% at 300s — deadline widening does nothing here, so the
remaining shortfall after the entries-limit fix is a second stable
cap, not settle time. Reverted the deadline back to 300s to match
loki_measure_pair rather than keep an unjustified change.

Wired the existing dump_loki_diagnostics helper (already used by
loki_measure_pair on a deadline miss) into loki_measure_frequency_pair
too — it's built around spec.logql + stats parsing, which is
query-shape-agnostic, so it works unmodified for the matrix path. If
L4 still falls short, the next run's failure carries the raw Loki
stats (chunk-fetch counts, any warnings) instead of another guess.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y
jensholdgaard added a commit that referenced this pull request Jul 17, 2026
* feat(bench): rfc 0031 l4 — wire into the live dispatch loop

Live-wires PairClass::L4 into rfc0031_indicative_comparative_run, the
#[ignore]d container-based dispatch test. The previous slice proved
the L4 machinery (ourios_aggregate_answer, parse_loki_matrix,
pick_frequency_pair, compare_aggregations) only at the fixture level,
against a hand-built Loki matrix response — this slice makes it real
against a running Loki container and the actual corpus.

L4 is picked and measured as its own step, kept OUT of the
`Picks`/`specs: Vec<PairSpec>` pipeline the L1/L2/L3/L6 classes share:
an aggregation's (bucket, group) -> count map is not a LineKey
multiset, and forcing it through OuriosAnswer/compare_lines would
misrepresent the state rather than model it (the same "make invalid
states unrepresentable" reasoning the miner/parquet layers already
follow). Concretely: pick_frequency_pair runs post-store-build like
pick_template_pair; its PairSpec is built with the exact dsl/logql
shape the fixture-level test already pinned; loki_query_matrix issues
a real query_range metric call with `step` pinned to the bucket
width so evaluation instants land on parse_loki_matrix's documented
bucket-alignment convention (t = bucket_start + width);
loki_measure_frequency_pair polls it to completeness the same way
loki_measure_pair does for line-returning pairs. Both share the same
Loki container and corpus replay as the existing pairs.

Equivalence-required-but-bytes-unasserted: RFC0031.1 (result-set
equivalence) is never optional, so run_l4_pair asserts
compare_aggregations(...).is_equal() unconditionally — an L4 mismatch
fails the run exactly like every other class's equivalence check. Only
the bytes RATIO stays unasserted (M_L4 is still §7-DEFERRED, no frozen
margin to gate against yet): print_l4_report reuses
print_pair_bytes_gates, which already prints L4's ratio with no
verdict. L4 is measured, equivalence-checked, and reported LAST — after
the L1-L3/L6 evidence has printed and their frozen gates have already
asserted — so an L4-only failure cannot destroy that evidence (the same
run #11 salvage lesson the rest of the harness follows). A missing
candidate is reported loudly at pick time, never silently skipped.

Purely additive: class_pair_specs, build_pair_specs, frozen_gate_failures,
print_pair_bytes_gates, print_indicative_report, PairSpec, and PairClass
are unchanged — no frozen-gate behavior for L1/L2/L3/L6 is touched.

Verification: cargo fmt --all --check, cargo clippy --all-targets
--all-features -- -D warnings (workspace), cargo nextest run -p
ourios-bench (165 passed, 7 skipped) and cargo test -p ourios-bench
--all-features all green, including the untouched fixture-level
rfc0031_5_l4_frequency_aggregation_bytes. The corpus-scale dispatch
test itself needs Docker + OURIOS_COMPARATIVE_CORPUS, neither available
in this sandbox — its first live proof is the comparative-bench
dispatch workflow, same as every other slice in this harness's history.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — backtick-delimit the loki regexp argument

The dispatch's first-ever run failed: capture_regex's own Go RE2
escapes (\s+, \S+) were embedded inside a double-quoted LogQL string
literal, which tried to interpret those backslashes as its own escape
sequences (\s is not a valid one) and Loki rejected the query with
"invalid char escape" before the pattern reached the regex engine.
Fixed by switching to a backtick-delimited (LogQL/Go raw string)
regexp argument, which passes the pattern through literally.

Extracted the duplicated PairSpec-construction block (present
independently in the fixture test and the live-wiring loop) into one
shared l4_pair_spec helper, closing the drift risk and centralizing
the fix. Added a backtick guard: a capture_regex containing a
backtick (regex_escape does not escape backticks) would prematurely
close the raw string, so the candidate is now rejected loudly instead
of emitting a malformed query.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — measure L4 before, not after, the L1-L6 failure asserts

The second dispatch run failed on the pre-existing, documented L3
Loki-side flake (0 of 9 rows before timeout) — but the run never even
attempted L4: the failures.is_empty() assert for L1-L6's own salvaged
measurement failures sat textually BEFORE the L4 measurement/report
code, so any earlier pair's failure aborted the test before L4 was
ever reached. This inverted the design intent (an L4-only failure
should not destroy L1-L6 evidence, not the other way around).

Moved L4's measurement to run immediately after the report prints,
before the gate/failures assertions. run_l4_pair now pushes a Loki-side
measurement failure (flake) into the same failures vec the other
classes salvage into, instead of panicking immediately — so a flaky
L4 measurement no longer aborts before the L1-L6 evidence is captured,
symmetric with the fix for the reverse direction. A genuine L4
equivalence MISMATCH still hard-panics immediately, unchanged:
RFC0031.1 equivalence is never optional, matching L1-L6's own
compare_lines assertion.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — cap the picker's row-count ceiling at 100K

The third dispatch got past the control-flow fix and genuinely
measured L4 — but the picked candidate (a service's dominant,
near-catch-all template) summed to ~971K matching rows, and Loki
returned only 811,775 of them before the 300s poll deadline (the
same budget every other class's loki_measure_pair uses). L4_MIN_ROWS
was a floor with no ceiling, so the picker had no reason to prefer a
smaller, still-meaningful candidate. Added L4_MAX_ROWS=100_000
(comfortable margin at the observed ~2.7K rows/s Loki throughput) to
frequency_shape_rejection, so the picker moves on to a candidate the
poll can actually finish measuring.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — disable loki's query-range results cache

Run #4's L4 pair plateaued at 11,053/11,523 rows across every 10s poll
instead of climbing to completeness. The picker's row ceiling (run #3's
fix) had already ruled out "too large to finish in time" — the count
never moved at all, which points at a cache serving the same stale
answer on every retry rather than a slow ingest.

Loki's bundled local-config.yaml enables the embedded results cache for
query_range's metric/matrix path (L4's loki_query_matrix), keyed by the
query+start+end+step tuple that loki_measure_frequency_pair repolls
unchanged. The first (still-incomplete) response gets cached and echoed
back on every subsequent poll. Plain log queries (loki_query_range, used
by L1-L3/L6) aren't extent-cached the same way, so they self-heal across
polls untouched by this.

-query-range.cache-results=false trades Loki's own query latency for
correctness of the harness's completeness poll — in Loki's favour, same
as the other operator-tuning flags already on this container.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — correct the results-cache disable flag name

Run #5 never got past container startup: `-query-range.cache-results`
doesn't exist ("flag provided but not defined"), so Loki's /ready check
timed out on a container that failed to start at all.

Checked the pinned v3.5.3 source directly instead of guessing again:
queryrangebase.Config.CacheResults is registered under the `querier.`
flag prefix in roundtrip.go, not `query-range.`. Correct flag is
-querier.cache-results=false.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — widen the loki poll deadline to 900s

Run #6 (with the corrected -querier.cache-results=false flag from the
prior commit) proved the results-cache theory wrong: L1-L3/L6 all
measured cleanly, but L4 still plateaued — 10752/11523 rows (93.3%),
even slightly worse than run #4's 95.9% pre-fix, and the shortfall
varies run to run rather than repeating a fixed cached answer.

That points at genuine, variable completion time rather than a bug:
L4's LogQL runs a `| regexp` capture over every candidate line before
grouping and counting, a real per-line cost the other classes' plain
stream/count queries never pay. Widened loki_measure_frequency_pair's
deadline from 300s to 900s — well inside the CI job's unset (360 min
default) timeout given the whole run has taken ~95-100 min so far.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — raise loki's max-entries-limit, revert deadline theory

Runs #4/#6/#7 all converged L4 to ~93-96% of expected rows, independent
of poll deadline (300s vs 900s made no measurable difference) — ruling
out both a results-cache echo (already disabled in #fe5915a) and a
"just needs more time" theory (the deadline widening from the prior
commit). A stable, time-independent shortfall points at something being
permanently excluded, not merely delayed.

Pulled the frozen otel-demo-v8 corpus locally and checked every log line
matching the L4 pair's needle ("Wrote producer snapshot at offset")
against its capture regex directly: all 11,525 matches parse cleanly.
The regex/content isn't the problem — some matching lines are never
being scanned at all.

That points at Loki's default -validation.max-entries-limit (5000):
count_over_time with a |regexp stage has to scan every raw kafka log
line in a query-frontend split before the line filter narrows it down,
and kafka's per-split volume exceeds 5000 lines often enough to
silently truncate the scan before every matching line is reached.
Raised the limit well past the corpus's noisiest single template's
volume (~971K rows).

Reverted the 900s deadline back to 300s (matching loki_measure_pair) —
the widened deadline never addressed the actual bottleneck, and keeping
it would misattribute the fix in a way that'd mislead the next reader.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — widen poll deadline now the entries cap is gone

Run #8 (max-entries-limit raised) moved L4 from a hard ~93% plateau to
97.1% (11192/11523) — real progress, and unlike runs #4/#6/#7 the
remaining gap now plausibly behaves like genuine ingest settle time
rather than a fixed ceiling, since the artificial cap that made the
prior 300s vs 900s test inconclusive is gone. Widened the deadline to
600s to test that directly before assuming a third factor is at play.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — revert unhelpful deadline widening, add diagnostics

Run #9 (600s) measured 96.5% (11123/11523), statistically the same as
run #8's 97.1% at 300s — deadline widening does nothing here, so the
remaining shortfall after the entries-limit fix is a second stable
cap, not settle time. Reverted the deadline back to 300s to match
loki_measure_pair rather than keep an unjustified change.

Wired the existing dump_loki_diagnostics helper (already used by
loki_measure_pair on a deadline miss) into loki_measure_frequency_pair
too — it's built around spec.logql + stats parsing, which is
query-shape-agnostic, so it works unmodified for the matrix path. If
L4 still falls short, the next run's failure carries the raw Loki
stats (chunk-fetch counts, any warnings) instead of another guess.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — epoch-align the loki query window to bucket boundaries

Run #10's diagnostics (wired in the prior commit) confirmed the L4
query itself is well-formed and Loki answers it successfully — no
error, no chunk-fetch shortfall visible in the sampled response. That,
combined with runs #6-#10 all converging to a stable ~93-97% regardless
of poll deadline (300s/600s/900s all statistically indistinguishable),
rules out both a timing race and a malformed query.

The real mismatch: Loki's query_range evaluates a step-grid starting
exactly at `start` (start, start+step, ..., end), but Ourios's own
bucket(width) semantics are epoch-aligned (floor(ts/width)*width) —
`min_effective_time_unix_nano` (the corpus's raw earliest timestamp)
has no reason to already be a bucket-boundary multiple. Unless
(end - start) is an exact multiple of the bucket width, the step-grid
leaves a fractional sliver at the tail of the range with no evaluated
window covering it at all — real, ingested, settled data that's simply
never queried, independent of poll duration. That's exactly the shape
every run has shown.

Snap `start` down and `end` up to the nearest bucket-width boundary in
l4_pair_spec — costs nothing (no data exists outside [min, max] to
inflate the count) and guarantees the step-grid fully covers the range.

Verified: cargo fmt --all --check, workspace cargo clippy -D warnings,
local (non-container) rfc0031_comparative unit tests — 39 passed.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* diag(bench): rfc 0031 l4 — add an ingest-vs-query-side split probe

Run #11 (bucket-aligned query window) measured 96.6% (11133/11523) —
narrower than the pre-fix ~93% plateau, but still in the same stable
band as runs #6-#10, all independent of poll deadline, entries-limit,
and now bucket alignment. Six straight dispatches without closing the
gap means continuing to guess at query-side LogQL/config knobs isn't
warranted anymore.

Added a decisive probe: on a deadline miss, loki_measure_frequency_pair
now also runs a PLAIN line-filter count (no count_over_time, no
regexp) for the same needle + window via the new
loki_query_range_uncapped (limit sized to expected_rows, unlike the
shared loki_query_range's fixed 5000 cap — which is below this pair's
11523 expected rows and would itself lie about the count). If that
plain count also falls short by the same margin, the shortfall is
ingest-side (Loki never stored those lines) and no further query
tuning will fix it; if it's ~complete, the loss is specific to the
aggregation path.

Diagnostic-only change — no behavior change to the measurement itself,
just evidence gathering on the existing failure path.

Verified: cargo fmt --all --check, workspace cargo clippy -D warnings,
local (non-container) rfc0031_comparative unit tests — 39 passed.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — reject high-frequency candidates prone to loki dedup

Run #12's decisive diagnostic confirmed the L4 shortfall is ingest-side:
a plain unaggregated line-filter count for the same needle+window came
back just as short (11160/11523) as every aggregation-path attempt.
Loki's ingester silently drops a log entry that collides with another
on (timestamp, body) within the same stream — a drop invisible to the
OTLP push response's partial_success (push_otlp already asserts that
field is clean on every push in every run so far). No query-side fix
was ever going to close this gap; the picker was choosing a candidate
Loki structurally can't ingest identically.

kafka's template_id=16 ("Wrote producer snapshot") fires roughly every
15s. A local exploration against the real frozen corpus (offline, no
Loki container — pick_frequency_pair only touches Ourios's own
pipeline) found candidates at much lower frequency clear of the same
floors: template_id=60 ("Periodic task") at ~144s average cadence, ~10x
the failing candidate's margin.

Added L4_MIN_AVG_INTERVAL_SECONDS (100s) to frequency_shape_rejection
as a durable picker rule, not a one-off override — this protects any
future re-run of the picker against landing on another collision-prone
high-frequency template, not just this specific dispatch.

Updated two pre-existing tests (pick_frequency_pair_finds_a_moderate_
cardinality_group, rfc0031_5_l4_frequency_aggregation_bytes) whose
synthetic sub-3s timelines — convenient for test speed, not meant to
model real timing risk — tripped the new floor; scaled their
timestamps 1000x (preserving cardinality/row-count/needle assertions
unchanged) so they represent a realistic, non-collision-prone example.

Verified: cargo fmt --all --check, workspace cargo clippy -D warnings,
local (non-container) rfc0031_comparative unit tests — 40 passed
(39 prior + 1 new: frequency_shape_rejection_enforces_the_average_
interval_floor).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* diag(bench): rfc 0031 l4 — dump loki's own container logs on a deadline miss

Run #13's lower-frequency candidate (template_id=60, ~144s average
cadence) still fell short (1144/1197), and a corpus-side check ruled
out the leading theory entirely: every one of the 1197 matching
records has a UNIQUE timestamp AND a unique body (verified via jq
against the frozen otel-demo-v8 corpus locally) — zero exact
(timestamp, body) collisions possible. Loki's documented dedup rule
cannot be the mechanism here, which means it likely wasn't the full
story for the prior candidate either, even though lowering the
frequency floor did measurably help (17.5% loss -> 4.4% loss).

Also checked push_corpus_to_loki/push_otlp end to end for a harness-
side drop: the batching loop appends every non-empty corpus line's
resource_logs to `pending` before any flush, with a final flush after
the read loop — no line is skippable, and push_otlp's retry resends
the identical Bytes payload, so no bug found there either.

Everything checkable from the client side (query responses, corpus
content, our own push code) is now ruled out or confirmed clean. The
next place to look is Loki itself: on a deadline miss,
loki_measure_frequency_pair now also dumps the Loki container's own
stderr, filtered to warn/error/drop/reject/rate-limit/stream-limit
lines — the ingester logs these for exactly the mechanisms still on
the table (rate limiting, out-of-order rejection, stream-limit drops),
none of which are visible in a query response or push_otlp's already-
clean partial_success check.

Diagnostic-only — no change to measurement behavior.

Verified: cargo fmt --all --check, workspace cargo clippy -D warnings,
local (non-container) rfc0031_comparative unit tests — 40 passed.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* diag(bench): rfc 0031 l4 — scrape loki's discarded-samples metrics

Run #14's level=warn/level=error stderr scan came back with zero
matches in 6618 total lines — whatever is causing the L4 shortfall
(still 1147/1197 with the lower-frequency candidate), Loki doesn't
consider it log-worthy. That rules out rate limiting, out-of-order
rejection, and stream-limit drops as commonly logged at WARN.

(The first attempt at the stderr filter was a naive "contains warn"
substring match, which drowned in false positives from query text
like `severity_text="WARN"` appearing inside level=info lines — fixed
to match on the `level=` field precisely.)

Loki's distributor increments loki_discarded_samples_total/
loki_discarded_bytes_total (labeled by reason) even for discards that
don't warrant a log line — its own dedicated counter for exactly this
question. Added dump_loki_discard_metrics, scraping /metrics on a
deadline miss (extracted as its own function, alongside
dump_loki_diagnostics, to keep loki_measure_frequency_pair under
clippy's line-count lint).

Diagnostic-only — no change to measurement behavior.

Verified: cargo fmt --all --check, workspace cargo clippy -D warnings,
local (non-container) rfc0031_comparative unit tests — 40 passed.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* feat(bench): rfc 0031 l4 — documented completeness margin (§7, 2026-07-17)

Sixteen dispatches (runs #1-#16) exhausted every mechanism checkable
from the harness's side without ever reaching exact L4 completeness.
Runs #13-#16, specifically, ruled out: query-shape artifacts (a plain
line-filter count matched the aggregation-path shortfall exactly),
Loki's documented same-(timestamp,body) dedup (zero exact collisions
found via direct corpus analysis), interleaving between a genuine
mid-corpus kafka restart's two service-instance periods (cleanly
sequential), a harness-side push/batching bug (read end to end, none
found), anything Loki logs at WARN/ERROR (zero matches bar one
harmless startup transient), and Loki's own discarded-samples
Prometheus accounting (zero discards of any kind, any reason).

This matches an open, unresolved upstream Loki issue
(grafana/loki#10658 and related): wide-time-range queries silently
missing a small, consistent percentage of lines, with no error, no
discard signal, and no maintainer-identified root cause. It's a
documented, external, currently-unfixable characteristic of the
comparison partner, not an Ourios or harness defect.

Adds L4_COMPLETENESS_MARGIN = 0.90 (real headroom over the observed
3.9-4.4% loss band) and compare_aggregations_within_margin — narrowly
scoped: it still hard-fails, at any margin, on Loki reporting MORE
than Ourios for any cell or a cell absent from Ourios's own answer,
the two signals that would actually indicate a correctness bug. Only
aggregate under-counting up to the margin is tolerated.
compare_aggregations (exact) is untouched and still gates the
RFC0031.5 fixture-level test's synthetic Loki answer.

Wires the margin into both loki_measure_frequency_pair's poll-complete
threshold (accept short-of-exact within margin instead of always
retrying to a hard timeout) and run_l4_pair's equivalence assertion.

RFC 0031 amended: RFC0031.1's L4 clause now states the margin
explicitly, and §7's L4-query-shape entry (previously open) is closed
with the full evidence trail and the margin decision. M_L4 (bytes-read)
stays deferred — this unblocks a measurement, it doesn't freeze that
margin.

Verified: cargo fmt --all --check, workspace cargo clippy -D warnings,
ourios-bench lib tests (32 passed, comparative module) + local
rfc0031_comparative integration tests (40 passed), mdbook build.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — margin check is total-level, not per-cell

Run #17 validated the completeness-margin design and immediately
refined it: the poll-completion check passed cleanly (1153/1197,
96.3%), but the equivalence check then hard-failed on a single cell
landing 1 row OVER Ourios's count (114 vs 113 for one bucket/value)
while the aggregate total stayed a solid under-count — consistent
with the same step-grid boundary imprecision already characterized
(a record landing in an adjacent bucket), not fabrication.

The original compare_aggregations_within_margin checked "Loki > Ourios"
per cell, which was too strict for that kind of noise. Refined to
check for phantom cells (a (bucket, group_key) Loki reports that
Ourios's own answer doesn't contain at all) and Loki's TOTAL exceeding
Ourios's total instead — this still catches the failure mode that
would actually indicate a bug (wrong regex or wrong bucket math would
produce cells Ourios never produced at all, or push the total over)
while tolerating single-cell boundary noise on keys both systems agree
exist.

Added a test for the exact run #17 shape (a single cell over, total
still under, must pass) alongside the existing phantom-cell and
net-overcount tests (renamed from "overcount" now that the check is
total-level). RFC 0031 §7's L4 entry updated with the refinement and
its rationale.

Verified: cargo fmt --all --check, workspace cargo clippy -D warnings,
ourios-bench lib tests (33 passed, +1 new) + local rfc0031_comparative
integration tests (40 passed), mdbook build.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — per-group_key margin, address PR #536 review

PR #536's code review (14 fresh findings across Copilot + CodeRabbit's
post-run-18 passes) surfaced one substantive correctness gap and
several real documentation/robustness issues in the completeness-
margin work. All verified against current code before fixing.

Substantive fix — cross-key redistribution gap (CodeRabbit, Major):
compare_aggregations_within_margin's grand-total-only check (from the
run #17 fix) let Loki over-count one group_key while under-counting
another by the same amount and still read as 100% complete: Ourios
{A: 100, B: 100} vs Loki {A: 190, B: 10} sums to a "complete" 200/200
while hiding A being fabricated to compensate for B being nearly
lost. Refined to aggregate ourios/loki BY group_key first (summing
each key across every bucket it appears in), then apply the
phantom/overcount/margin checks per-key. This still tolerates run
#17's exact shape (a single bucket's +1 doesn't change a key's own
total across its buckets) while rejecting the redistribution a pure
grand-total check missed. Added regression tests for both shapes.

Also populates real per-key examples in mismatch reports (Copilot: the
old design returned examples: Vec::new() on both mismatch paths,
despite the function accepting examples_cap and RFC0031.1 calling for
example keys on a failed comparison).

Documentation/robustness fixes (all verified against current code,
none required a runtime-behavior change beyond the fix above):
- Two stale comments still asserted Loki's same-(timestamp, body)
  ingester dedup as the shortfall's mechanism, contradicting the
  nearby docs that say this was directly disproven and the true
  mechanism is uncharacterized (Copilot, 6 threads pointing at 2 real
  sites: the frequency_shape_rejection rejection message and one test
  comment — the other 4 threads were already-accurate historical
  narrative, verified and left alone).
- Missing `//` justification comment on one #[allow(cast_precision_loss)]
  (CodeRabbit).
- L4 picker silently continues when no viable candidate exists
  (l4_spec.is_none()) — only an eprintln, no run failure, despite L4
  being a must-win class (CodeRabbit). Now pushes into `failures`.
- PR description inaccurately described L4's equivalence assertion
  (exact compare_aggregations, ordered after the frozen gates) —
  rewritten to match actual behavior (margin-based, before the frozen
  gates, matching the run #11 salvage design already documented inline).
- RFC 0031 §7's L4 entry claimed the picker "prefers the lowest-
  frequency viable candidate" — the actual algorithm is first-fit in
  ascending (template_id, param) order, not an exhaustive ranking
  (CodeRabbit, tagged Heavy Lift). Reworded to describe actual
  behavior and the deliberate scope decision (a real ranking pass
  would cost a query per candidate against a corpus with tens of
  thousands of templates; not built given first-fit has now found a
  validated candidate three real dispatches running).
- Double-backtick delimiters for a LogQL code span containing literal
  backticks (CodeRabbit, MD038).

Verified: cargo fmt --all --check, workspace cargo clippy -D warnings,
ourios-bench lib tests (34 passed, +2 new) + local rfc0031_comparative
integration tests (40 passed), mdbook build.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — absolute row tolerance, not pure percentage

Run #19 (the verification dispatch for the round-1 review fixes) found
a real edge case in the per-group_key percentage margin: a group_key
with exactly 1 total Ourios row, where Loki captured 0 (0%). A pure
ratio has no meaningful middle ground at n=1 — it's binary, 0% or
100% — yet losing one isolated occurrence is fully consistent with
the already-characterized ~4-8% aggregate loss rate this whole margin
exists to tolerate.

Converted the per-key check from a ratio (loki/ourios >= margin) to an
absolute row tolerance floored at 1: ceil(ourios_key_total *
(1 - margin)).max(1). This tolerates a cardinality-1 key losing its
only row while still catching a real shortfall on a large key (100
rows, tolerance 10, losing 20 still rejects) — the phantom-cell and
per-key-overcount hard-fail checks are unaffected. Two new regression
tests cover both ends.

Verified: cargo fmt --all --check, workspace cargo clippy -D warnings,
ourios-bench lib tests (36 passed, +2 new) + local rfc0031_comparative
integration tests (40 passed), mdbook build.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — margin comparator precision, review triage

compare_aggregations_within_margin's tolerance formula (ceil(o*(1-margin))
.max(1)) was itself miscalibrated for small-but-not-1 totals, per two
independent Copilot review threads (o=2 at 90%: tolerance=1 permits 50%
completeness, not 90%). Replace the subtract-then-round row tolerance
with a direct, epsilon-guarded comparison — loki_total >= ourios_total *
margin — which also sidesteps a second bug the naive floor() fix
introduced: 1.0 - 0.9 isn't exactly 0.1 in f64, so floor(40.0 * (1.0 -
0.9)) truncated to 3 instead of 4, tightening the tolerance at exact
90%-boundary cases (caught by the existing
margin_comparison_tolerates_undercount_within_margin test's svcB case).

Extract phantom_cells and aggregate_by_group_key helpers to bring the
function back under clippy's line limit, and add a # Panics section for
the margin-validation assert.

Fix several accumulated PR #536 review findings: run_l4_pair's doc
comment claimed L4 runs after the L1-L3/L6 frozen gates assert (it
actually runs before, printing first); three "ingest-vs-query" overclaims
(the plain line-filter probe still calls query_range, so it can rule out
"specific to the metric-aggregation path" but not prove ingest-side
loss); L4_COMPLETENESS_MARGIN's own doc comment still described the
superseded total-level design. Add a clarifying comment on l4_pair_spec's
step-grid reasoning (the first evaluated instant decodes to an empty
phantom bucket, not a lost real one) and fix the RFC's LogQL code span,
which kept the backslash escaping needed for single backticks even after
switching to a double-backtick delimiter that makes it unnecessary.
Reconcile RFC0031.5's must-win predicate with M_L4 staying deferred —
add a note that the predicate is the target contract, not currently
gated.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — panic-safe diagnostic probe, more review triage

loki_query_range_uncapped used expect()/assert!() internally, but it
runs on the L4 deadline-miss diagnostic path inside the same
runtime.block_on that gathers L1-L3/L6's evidence — a panic there
(a real Loki error response, a malformed body) would unwind the whole
async block and lose all of it, defeating the print-before-assert
salvage design (Copilot). Converted to return Result<u64, String>
instead of panicking, matching the already-panic-free sibling
diagnostics (dump_loki_diagnostics et al.).

Also: fix a test comment that said "one row under the ceiling" for a
fixture that actually lands exactly at the ceiling; fix an unreachable!
message's imprecise invariant claim (the real gating condition is
l4_spec.is_some() implies l4_loki.is_some(), not "iff frequency is
Some"); document loki_query_matrix's whole-second/bucket-alignment
precondition and verify it against l4_pair_spec, its only caller;
reorder loki_measure_frequency_pair's deadline-miss diagnostics to run
only when the completeness margin is actually missed, not on every
deadline-miss regardless of outcome; fix two fixture comments claiming
"~300s average spacing" that don't match their own timestamps (actually
~580s) and a comment attributing the L4 shortfall to ingest-side dedup
after that theory was directly disproven elsewhere in the same file.

Verified the remaining ~40 accumulated review threads (mostly a
recurring "shared 300s deadline" doc/code mismatch and the run_l4_pair
ordering claim, duplicated across many review rounds) against current
code: all already correct, superseded by earlier commits in this
investigation.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — round-4 review triage on the margin comparator

Validate margin at function entry rather than after the phantom/overcount
checks, so an invalid margin always panics per the documented contract
instead of potentially returning a data-shaped mismatch first
(CodeRabbit). Fix the doc comment paragraph still describing the
superseded floor-based tolerance (Copilot). Reword the L4-skip
diagnostic and failure message to name both reasons l4_spec can be None
(picker bounds vs a backtick in the capture regex) and to make clear the
skip fails the dispatch rather than reading as benign (Copilot, two
sites).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* fix(bench): rfc 0031 l4 — margin=1.0 strictness + panic-safe matrix poll

Two Copilot findings on the previous commit, both verified genuine:

The cardinality-1 exemption applied at any margin, so a caller passing
margin = 1.0 (exact completeness) would still accept Loki returning 0 of
1 for an n=1 key — the exemption now only applies to a genuinely
fractional margin, with a regression test covering both directions at
1.0. Bit-identical behavior at the harness's 0.90.

loki_query_matrix still used expect/assert internally, so a transient
transport error, 5xx, or torn body during the L4 poll — which runs LAST
in the same async block holding every other pair's already-collected
measurement — would panic and unwind all of it. Converted to
Result<L4Measured, String>; the poll loop now retries an Err until its
deadline exactly like an incomplete answer, then surfaces it as the
pair's failure. Extracted the below-margin shortfall diagnostics into
dump_l4_shortfall_diagnostics to stay under clippy's function-length
limit.

Neither change alters the measured semantics run #23 is currently
confirming (the comparator formula is untouched; at margin 0.90 the
exemption gating is unchanged).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* docs(bench): rfc 0031 l4 — document why the phantom check is cell-level

Copilot's latest pass proposed weakening phantom detection from
(bucket, group_key) cells to bare group_keys so a boundary-exact record
shifting into an empty adjacent bucket can't read as phantom. Declined:
a systematic bucket-decode error (every cell shifted one width — the
run #11 bug class) leaves every per-key total intact, so the cell-level
check is the only guard that catches it, while the false positive it
risks requires a nanosecond-exact bucket-boundary timestamp that no
real dispatch has ever produced. Documented the trade-off on
phantom_cells instead of changing behavior.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants