fix: stabilize extension turns, transcript hydration, and reasoning controls

[effortprogrammer]

Summary

Fixes #1329, #1076, and #891. Addresses the numeric-counter near-repeat case in #1330 with progress and code-safety guards.

• Run and apply before_agent_start for idle custom-trigger turns. Preserve final admission, cancel held hooks without a ghost provider call, and wait for deferred recovery before print-mode output.
• Separate warmed transcript caches from visible history. Incomplete renders keep their initial boundary; the complete prefix is published once hydration finishes.
• Detect bounded numeric paragraph drift while preserving tool progress, identifiers, lists, and nested fenced examples. This is intentionally not semantic paraphrase detection.
• Give retained affected built-in models explicit reasoning maps matching their implemented controls, preserve graded models, and remove the stale Xiaomi disabled-thinking override.
• Add provider-scoped generation with own-key selector validation so unrelated provider data remains byte-identical.

The fork was synchronized with upstream 89a06b08, and that commit is merged here as 758ee7a0. Upstream-retired Moonshot models were not restored: the affected capability table now covers 52 retained entries, with explicit tests for the retired rows.

Verification

• bun run check and the complete workspace build passed on the latest integration.
• Affected integration regressions: 249 passed across AI, coding-agent, and codemode.
• Retired-catalog contrast: the old catalog exposes the 12 retired provider/model pairs; the current catalog exposes none.
• Real source CLI/RPC hook QA: 13/13.
• Real source CLI numeric-loop recovery and code/progress controls: 13/13.
• Real native-PTY transcript resume: 80-to-120 and 120-to-80 passed, with an observed incomplete render retaining the same history boundary.
• Real source RPC reasoning/wire QA: 15/15.
• Real native-PTY reasoning commands, persistence, and restart QA: 13/13.
• Five fresh terminal captures were replayed through actual Chrome/xterm.js and visually inspected. Captures use PTY output events, not keyboard-input substitutions.
• All QA used isolated local providers and sandboxes. No real model credentials or paid provider calls were used; PTYs, servers, browsers, and temporary state were cleaned up.

One local full-workspace run on macOS/Node 26 recorded failures in 12 coding-agent files, dominated by test/hook timeouts and a child-process completion failure. The same 12 files then passed 109 tests in an isolated single-worker run on unchanged code, without increasing timeouts or skipping tests. The broad-run log is retained; it is not being presented as a full-suite pass. Linux/Node 24 PR CI is a separate verification gate.

Reproducible drivers are included under .agents/skills/senpi-qa/scripts/scenarios/:

node .agents/skills/senpi-qa/scripts/scenarios/issue-1329-qa.mjs --evidence issue-1329
node .agents/skills/senpi-qa/scripts/scenarios/issue-1330-qa.mjs --evidence issue-1330
node .agents/skills/senpi-qa/scripts/scenarios/issue-1076-qa.mjs --evidence issue-1076 --self-test
node .agents/skills/senpi-qa/scripts/scenarios/issue-891-qa.mjs --evidence issue-891
node .agents/skills/senpi-qa/scripts/scenarios/issue-891-tui-qa.mjs --evidence issue-891-tui

The last driver requires Bun >=1.4 on PATH (or selected through SENPI_QA_BUN) and the skill-local node-pty dependency.

Maintainer note

CONTRIBUTING.md reserves CHANGELOG.md entries for maintainers, so no contributor changelog edit is included. Exact nearest changes.md coverage is present. The runtime-change changelog gate needs the corresponding maintainer entry or explicit maintainer policy handling; this user-facing change is not being mislabeled as no-changelog.

Capability QA proves Senpi's catalog, UI, persistence, and request controls. It does not claim live account entitlement or provider availability for every retained model.

---

Summary by cubic

Fixes #1329, #1076, #891, and the numeric near-repeat case in #1330. Idle custom-trigger turns now run before_agent_start, print mode waits for recovery turns, and hydrated transcript history publishes atomically instead of exposing partial prefixes.

Behavior changes

• Idle send_custom_message turns apply admission hooks and cancel without a ghost provider call.
• Print mode waits for settled session work before selecting the final result.
• Progressive transcript keeps its initial visible boundary until full history is warmed; unrelated renders and resizes no longer show partial history.
• Numeric paragraph drift is detected with bounded normalization, while code fences and tool progress are preserved; this is not semantic paraphrase detection.
• Reasoning maps now match each built-in adapter's controls, graded models stay graded, and Xiaomi's disabled-thinking support is no longer overridden.
• Qwen3.8 Token Plan rows now send the documented graded effort values instead of clamped toggle-only controls.
• generate-models.ts supports --providers for provider-scoped regeneration, leaving unselected files byte-identical.
• The fork was re-synced with upstream main, bringing in overflow/SDK recovery, Astra context caps, and the RPC watchdog ownership repair; conflicting OpenCode data was regenerated in the process.
• Upstream-retired Moonshot models remain removed from the catalog.

Verification and rollout

• Added five reproducible QA scripts under .agents/skills/senpi-qa/scripts/scenarios/ plus regression tests in packages/ai and packages/coding-agent.
• Inspector resume handoff tests now close each debugger only after its own resume acknowledgement.
• QA used isolated local providers; no real credentials or paid provider calls.
• Unreleased changelog entries record the reasoning controls, extension-turn hooks, and atomic transcript hydration; changes.md coverage is included.

^{Written for commit 762d19a. Summary will update on new commits.}