Close max_audit.md in full — M7, the LOW table, and the roadmap

[ojassug]

Closes max_audit.md. Supersedes #23, which was closed as a side effect of renaming this branch — it covered only the first of these three commits.

commit	what
219b029	M7 — the Gateway forwards the caller's bytes, and measures them (DECISIONS §54)
d006cf7	L1, L4–L9 — the LOW table was never scheduled (DECISIONS §55)
a80e32d	ROADMAP — v1.5.0 is the baseline, and the audit is closed in full

---

The audit was not closed, twice

max_audit.md was declared closed while M7 was open, then declared closed again while the entire nine-row LOW table was open. Both were missed the same way: they sat in no wave table, and a check that enumerates the work that was done cannot see a finding nobody scheduled.

§54 named the rule after the first miss — an item that is in no table reads as done, exactly like a check that never ran reads as a pass — and it recurred three days later, which says the rule was not the fix. The fix is procedural and is now the first line of status doc §8: close a document against its own list of findings, not against the list of work that was done.

L2 and L3 were already closed, incidentally, by the C2 Buffer work. That is the tell — the two that got fixed are the two that happened to sit inside someone else's diff.

M7 (219b029)

When an elision fired, the proxy rebuilt the request with JSON.stringify, rewriting fields it had never touched:

client sent	provider received
"temperature": 1.0	"temperature":1
"top_p": 1e3	"top_p":1000
"seed": 12345678901234567890	"seed":12345678901234567000

The first two are cosmetic. The third is a different number — an integer past 2^53 does not survive JSON.parse → JSON.stringify — so a provider was asked for a seed the caller never chose, by a proxy whose whole promise is faithfulness. Elided content is now spliced into the caller's own bytes; where the caller's escaping differs from ours the splice declines and the original body goes out, because a lost saving costs tokens and a corrupted field costs correctness. A forwarded body can no longer be larger than the one that arrived (M7's third consequence, never previously asserted).

Savings now measure the bytes forwarded rather than the bundle render: 48.5% claimed against 47.1% actual, now 46.3% against 46.5%.

The LOW table (d006cf7)

Fixed: L1 (TOKENDAMPER_* enum values rejected rather than dropped), L7, L8 (escaped newlines counted). Recorded at their sites as acceptable rather than correct: L4, L5, L9. Comment corrected: L6 named a branch-and-bound search the code does not perform.

L7 was rated too low

The audit calls it "fails safe (skip) but silently loses the region". Measured end-to-end it loses the whole file, because a one-region file has nothing else to elide:

	in	out
no blank line after def	434 B	96 B (77.9%)
blank line after def	436 B	436 B (0%, fallback)

indentOf('') is 0, so the region began at column 0, the marker inherited column 0, and PythonValidator rejected it. The last scan in the same function already skipped blank lines; only the line reading the indent did not.

L1 is §30 arriving by the other door

TOKENDAMPER_PLANNER_MODE=session_dedup was silently discarded while --planner-mode session_dedup threw — and session_dedup is a real member of OptimizationMode, so a user had every reason to think it applied. All four enum parsers now reject through one helper rather than one being fixed and three keeping the trap.

Important

Technically breaking, for anyone with a typo'd TOKENDAMPER_* variable set — it used to be ignored and now errors. Accepted sets are unchanged.

Measurement

576 of 576 corpus rows byte-identical across all fifteen compared fields, per-row over one frozen 288-file corpus, both routes.

Warning

That does not mean L7 is inert. 0 of 45 Python corpus files have a blank line after a def — pip internals and this repo's own Python are uniformly PEP 8 there. The instrument is blind to this shape. This is §52's corpus-bias caveat with the sign reversed, and reading 576/576 as "no effect" is the mistake it exists to prevent.

Every fixed case was run against the unfixed engine first: 5 of 7 assertions fail there, and the 2 that pass are the negative controls, which must pass both ways or they are testing nothing.

669 tests, typecheck, lint and build green.

What this unblocks

The feature gate is open and the roadmap is current. Widening elision beyond TypeScript/JavaScript/Python is the one live item — every other corpus bucket is 0.00%. It is not one gate despite supportsRegionElision being one function: a new language needs a validator and extractSymbols coverage as well as a region scanner, and adding the scanner alone converts a 0% into a fallback via §33's measurement gate.

No feature work is started here.

🤖 Generated with Claude Code