feat(reject-gate + config): narrow gate, configurable model + reject_pattern via config.json

[fdaviddpt]

Update — config.json wiring (latest commits)

Beyond the reject-gate narrowing below, both knobs are now surfaced in config.json (the source of truth), with the env vars kept as overrides — per the "everything in JSON if possible" preference. model is per-project, so config is its natural home.

• log.sh (sourced by every script, incl. save-session.sh + run-consolidation.sh) bridges config.json → model / reject_pattern to REMEMBER_MODEL / REMEMBER_REJECT_PATTERN. Precedence via ${VAR:=...}: explicit shell env > config.json > built-in default. No Python change — haiku.py already reads these env vars.
• config.example.json: model + reject_pattern keys with _comments.
• README: both env rows note config.json is the source of truth.
• Tests: TestModelConfigBridge sources log.sh and asserts default / from-config / env-overrides-config. 66 passed (test_haiku + test_layered_config).

Also documents REMEMBER_MODEL in the README env table, which #112 added but didn't document.

---

Context

Follow-up to #112 (merged). The REMEMBER_MODEL knob from that PR is great and untouched here. This corrects one risk in the merged reject-gate: the refusal pattern was broad enough to silently drop legitimate hedged summaries from the memory layer.

The intent of the gate is right — a model refusal was once stored verbatim as a memory entry, and that must never reach the memory layer. But the merged pattern also matched normal summary openers:

output	merged #112	this PR
There are no blockers; merged !24648	dropped	kept
Unfortunately the build broke; fixed in X	dropped	kept
It seems the cache was stale; cleared it	dropped	kept
I notice the staging DB drifted; resynced	dropped	kept
I cannot invent timestamps. Do you want to:	rejected	rejected

A visible junk entry is deletable. A silently-dropped real entry is invisible — the worse failure of the two.

Changes

• Narrow DEFAULT_REJECT_PATTERN to unambiguous refusal/clarification stems: i cannot|can't|won't|am unable, could you, please provide|paste|share, i'm sorry. The dangerous narrative openers (there are no, it seems|looks like|appears, unfortunately, i notice, bare sorry) are removed from the default. All five of feat: configurable model via REMEMBER_MODEL + refusal reject-gate #112's original refusal cases still reject.
• Add REMEMBER_REJECT_PATTERN env knob (mirrors REMEMBER_MODEL / REMEMBER_MAX_TURNS): blank → narrow default, none → gate disabled, anything else → custom case-insensitive regex. An invalid custom regex falls back to the default rather than crashing the backgrounded consolidation run.

This also keeps the gate's content-judgment opt-out available, addressing the "format validation is the shell's job" intent behind test_parse_response_headerless_summary / _raw_conversation_echo: set REMEMBER_REJECT_PATTERN=none and only the literal SKIP contract applies.

Tests

Added to tests/test_haiku.py:

• 5 regression guards for the hedged-summary false positives above.
• _resolve_reject_pattern default / blank-fallback / none-disables / custom / invalid-fallback.
• gate-disabled keeps a refusal; custom pattern applies.

tests/test_haiku.py: 56 passed. All of #112's original refusal + keep-real-summary cases remain green.