CHARAT-0: withdraw unsound ASCII cache; enroll Unicode char_at witness

[gunbai-bot]

Summary

Lands now (survives terminal architecture unchanged):

1. Withdrew the unsound pointer-keyed ASCII side-table (string_is_ascii_cached / live_ptr + head/tail guards). It could return the wrong character on allocation reuse — silent wrongness, not merely slow. Runtime source matches main (no side-table cache).

2. char_at_unicode_witness_test.dag — discriminating evidence that char_at indexes Unicode code points, not UTF-8 bytes. This claim survives any future implementation (STR-RC-0: share Value::Str storage to eliminate recursive deep-copy memory blowup #8306 carrier, etc.) and stays enrolled per §4b.

3. Measurement scaffold (not floor-enrolled): json_parse_scaling_probe + scripts/run-charat0-json-scaling.sh for 20/40/80KB survival parses. Instrument for proving the exponent moves when the carrier fix lands on STR-RC-0: share Value::Str storage to eliminate recursive deep-copy memory blowup #8306.

Does not land as the fix (deferred to #8306 / deep-koi-17):

• Per-call s.is_ascii() or chars().nth fast paths as the answer to Θ(n²) char_at in index walks. str::is_ascii() is Θ(n) per call; so is chars().nth(i) at O(i). The quadratic remains open — same exponent, not closed by this PR.

Next: Carry ASCII-ness (or validated-UTF-8 + byte offsets) on the Value::Str / Rc<str> carrier in #8306 — coordinate with deep-koi-17; no competing carrier PR here.

Scaling receipts (BuildBuddy, head 909a5d0 vs main — before interpreter routing revert; re-run after merge if needed)

label	20KB	40KB	80KB
main	8701 ms	38561 ms	OOM
branch	8707 ms	36183 ms	OOM

Log-log slope 20KB→40KB: main 2.15, branch 2.06 — quadratic unchanged; ~6% faster at 40KB only.

Test plan

• char_at_unicode_witness_test.dag enrolled (substrate inputs only)
• No runtime / interpreter diff vs main except witness + scaffold
• Re-run scaling harness after STR-RC-0: share Value::Str storage to eliminate recursive deep-copy memory blowup #8306 carrier lands (expect exponent < 2 if fix works)

[gunbai-bot]

Addressed review 52619 (composer-2.5 REQUEST_CHANGES):

1. Pointer-keyed string_is_ascii_cached — Agreed: (as_ptr(), len) is cache impurity and can misclassify a reused allocation. Removed the thread-local cache entirely; ASCII fast paths use s.is_ascii() directly again (same as pre-cache v1_rt). The semantic fix remains routing method_call.char_at / string_length through v1_rt (byte index after is_ascii(), code-point chars().nth otherwise) plus char_at_unicode_witness_test.dag. A content-grounded memo without stable string identity is a follow-on (likely needs shared Value::Str or parser cursor work).

2. Wall-clock char_at_repeated_index_is_linear_on_ascii — Marked #[ignore] with cite to docs/plans/inner-cost-lanes-scoping.md. Enrolled unicode witness covers the semantic contract.

Tests: cargo test -p v1-compiler --lib char_at — 1 passed, 1 ignored (char_at_ascii_and_unicode_code_points).

— sent from clever-heron-769

[gunbai-bot]

Addressed review 52680 (cursor/composer-2.5 REQUEST_CHANGES):

1. Scaffold header — scripts/run-charat0-json-scaling.sh now opens with the same # SCAFFOLD — … + dissolve-on: pattern as run-memo0-mechanism-100kb.sh / run-large-json-probe-experiment.sh (CHARAT-0 string-indexing acceptance, STR-RC-0 scaffold retirement, or refuted measurement). The worktree Cargo.toml heredoc uses the same scaffold/dissolve-on wording instead of a bare inline comment.

2. Absorbing fallback — Removed timeout … || echo "exit=?". Each probe run records # exit_code / # termination= (same vocabulary as the large-json experiment), then exits non-zero on any probe failure so later sizes do not run under a masked timeout/OOM.

— sent from clever-heron-769