Audit fixes P0-P3: correctness, honest tools, e2e rig, thinking dials

[mike-diff]

What

A full audit pass over the repo, every finding verified empirically first
(the real binary, headless, against a scripted openai+anthropic endpoint
with wire captures), then fixed in four waves. Unit tests were written
before each fix and proven by mutation: every breaker was applied and its
test confirmed red before being reverted.

The four commits

Commit	Wave	Contents
fix(p0)	correctness	write preserves file mode (a 0755 script kept its exec bit; a 0600 file was widened to 0644 on every overwrite). Anthropic adapter never serializes an empty assistant turn as content:null (a 400 on every later call in that session). max_tokens becomes a dial (profile max_tokens + max_output_tokens tuning) and a 400 naming the model's real output cap self-heals to it: 16000 -> 400 -> retry at 8192 verified on the wire, so sub-16k-output models work out of the box. A failed turn keeps exchanges whose tool results already changed disk (a completed write used to vanish from history while the file stayed). An unparseable judge verdict gets one JSON-only repair retry, then defaults to not-done instead of stopping the drive; transport failures still stop.
feat(p1)	honest tools	read gains offset/limit (1-based lines) and an over-cap read ends with [N of M lines shown; re-read with offset N+1] instead of a generic truncation: the model can finally reach the end of a big file. search names files over the 1MB guard it did not search (a silent skip let the model confidently report the wrong file set). loc counts mainstream extensions (java, rb, cpp, tsx, ...) and discloses what it did not count. The gate mod executable re-resolves per mutating call, so a boundary installed mid-session binds immediately. mcp_timeout_secs dial replaces the fixed 60s call ceiling.
feat(p2)	infra	Startup session scans (lastBrain, latestSession) read each file's header only, cut at the turns key: startup cost is O(header bytes) across every saved session, not O(all transcripts). A standing end-to-end rig (SESH_E2E=1, wired into CI) builds the binary and drives it headless against a scripted endpoint with isolated HOMEs, pinning all the regressions above plus the gate-mod path and the anthropic wire shapes.
feat(p3)	features	thinking_budget (anthropic extended thinking; max_tokens is raised to clear the budget) and reasoning_effort (openai hint; a max_tokens cap then travels as max_completion_tokens) as profile fields, both default-off. The adapter captures streamed thinking blocks with signatures onto the assistant turn and passes them back on the final assistant message of the next request: the API rejects tool-use continuations under thinking without that round-trip.

Validation

• Red-first tests; every fix breaker-verified by mutation (applied, test
  failed, reverted). Two of my own assertions and one pre-existing pin
  (which had codified the silent search skip) were corrected in the process.
• User-level runs per wave: sesh -p -yes against the scripted mock with
  wire dumps (9/9 original finding scenarios pass on the final tree).
• The e2e rig itself was proven source-sensitive: reverting a fix at the
  source makes the rig red.
• gofmt -l . clean, go vet ./... clean, go test ./... green; CI now
  also runs the e2e rig on every push and PR.

Behavior notes

• search and loc outputs can grow one disclosure line each; the old
  silent behavior was the bug.
• A failed or cancelled turn now leaves completed tool work in history (with
  an adjusted notice); only an unconsumed opening rolls back.
• New profile dials: max_tokens, thinking_budget, reasoning_effort;
  new tuning dials: max_output_tokens, mcp_timeout_secs (documented in
  the scaffolded tuning.json.example and ~/.sesh README).

Left open

• Live cost/success bench for thinking_budget against a real thinking
  model before recommending it beyond default-off (wire behavior is pinned
  by unit + e2e tests).