Local, telemetry-free log of "couldn't do this" signals for roadmap feedback

Scoped-down, safe version of the "agents self-heal by extending their own tools when something's missing" pattern from harness-reliability research (the browser-use "Bitter Lesson" piece) — not literal self-modifying code, which is out of scope for a local read-only tool, but the useful half of the idea: **structured signal about where the harness hit a wall.**

**Mechanism.** When a run ends in an explicit decline ("the files can't answer this"), a `hard_fail` that never resolves, or repeated tool errors on the same case, append a one-line structured record (question shape / tier-like tag, not the user's actual data) to a local, never-transmitted log file. Purely local — no phone-home, no telemetry service, nothing that leaves the machine; it's a file the user (or, for the maintainer's own dogfooding, the maintainer) can choose to look at.

**Why:** this is real product-feedback signal about where Fella's ceiling actually is in everyday use, distinct from and complementary to the benchmark-derived cases in #68 (folder-QA v2) — those are hand-authored guesses at hard cases; this is what real sessions actually hit.

Needs a decision on exactly what's safe to log (case shape only, never file contents or the literal question) before implementation.

The issue names no files, tests, or entry points. Start by mapping where explicit declines, unresolved hard_fail results, and repeated tool errors are finalized, then document a safe case-shape-only record and local file location. Done means the logging policy is decided and records never include literal questions, file contents, or transmitted data.

Written by the indexing model from the issue text.