Mark 0.2.109 in the changelog and ignore scratch dirs in lint

[TheGreatAxios]

Two changes, both prerequisites for cutting 0.2.109.

1. Changelog. Renames ## [Unreleased] to ## [0.2.109] - 2026-08-24; the release script requires a clean working tree.

2. Lint ignores for gitignored scratch dirs. Running bun run check locally on main produced 247,804 errors — none of them in repo source. eslint.config.js and .prettierignore ignored scratch/, .scratch/, .worktrees/, vendor/, dist/, and node_modules/, but not .claude/, tmp/, or .tmp/. All three are gitignored, so CI never had them and never saw this; locally .claude/ alone is 7.3 GB across 110 accumulated agent worktrees, each a full checkout with TypeScript sources.

After ignoring all three: 0 errors, 5324 tests pass. The gate is usable locally again.

0.2.109 contents, all already merged

Every item is a deletion or consolidation — nothing additive.

PR	What went away
#612	repetition.ts, period-detection.ts, tool-fingerprint cycle detection, the 100-turn backstop, leaf no-progress counter — net −3018
#613	HARD_BLOCK_SALVAGES, never-edited/never-acted/no-ship, requireEdit plumbing, the parent salvage-nudge subsystem — net −511
#608	duplicate kill ring, duplicate mapReactorLike, duplicated timeout race → src/util/budget-race.ts — −119 excl. tests
#610	prose classifiers and parent hints replaced by structured stop reasons
#607	default 30-turn leaf cap; two orchestrator prompt rules that enforced nothing
#606	AgentProfile.fleetTier, declared but unreachable from any config format
#603	test state leaking into the real ~/.corbits/projects

What is deliberately kept

Provider stream error handling, connection retry and backoff, the turn-budget stop, the director's soft check-in nudge, requireEvidence/incomplete-report, the hasReportEnvelope self-check, and TUI stall recovery. concurrent-lane-overlap still logs and still never blocks.

Known tradeoff

With the detection layer removed, a genuinely stuck run is no longer cut short — it surfaces as a burned context window rather than a clean stop, and the operator interrupts instead. That is the deliberate direction: the detectors were terminating healthy runs, including one measured lane with 59 assistant turns, 58 of them carrying tool calls, and a clean final report.