v2.15.0 - The Honest-Reporting Release

Four of the five defects in this release are the same mistake at different altitudes: the code said something that was not true — about which backend ran, about how long a request had, or about what would happen when a lane failed.

The headline was found by probing, not reading. With every response truncated, generate_file made three requests — all to local — and then reported backend_used: "nvidia_glm", a lane that was never called. The escalation flipped the variable on its final iteration and the loop then terminated. There was never any cloud spend; there was a false name in the response, and the same false name went into recordExecution and therefore into get_analytics.

Fixed

• Retry timeouts were frozen at attempt 1. generate_file and modify_file computed a budget once, before their retry loop, then reused it while scaling tokens ×1.5 per retry and sometimes switching lanes. Attempt 3 could request 2.25× the tokens of attempt 1 under attempt 1's deadline, so a budget that was too small surfaced as "the backend timed out". Each attempt is now sized from its own live token count and backend.
• Truncation reported a backend that never ran. The dead cloud-escalation branch is removed from both handlers. A truncation now stays on the lane the caller got, retries there with a larger budget, and reports honestly if it still does not fit.
• modify_file's truncation retry never executed. Its loop condition re-tested a value that attempt 1 had already assigned, so token-scaling retries were dead code and a truncated modify got exactly one attempt.
• modify_file attributed results to the wrong lane. Five of its six reporting and metric sites named the backend chosen before the retry loop rather than the one that served the request.
• Eight hardcoded backend references are gone from the two handlers. An operator who had configured only, say, Groq was previously pointed at a lane they never set up.

Changed

• A backend you name explicitly no longer cascades. It gets one attempt, then an error distinguishing "your lane was tried and failed" from "your lane could not be attempted at all". A backend the router resolved — from auto, a routing rule, or the content-length hint — cascades exactly as before. The API keys are yours, and silently rerouting a request you pinned to one lane can spend your credit on lanes you never asked for. Pass backend: "auto" for the previous behaviour.
• A computed timeout is now capped by the lane's declared config.timeout. On the shipped configuration a scaled modify_file retry computed 367500ms for a lane declared at 30000ms. The declared value outranks the handler's floor, so a lane declared below what a large request needs now fails fast rather than running past your declared patience. A lane that declares nothing is unchanged. If large generations or modifications on one backend start failing quickly after upgrading, raise that lane's declared timeout first.

Notes

• 17 tools, unchanged. Nothing here adds, removes, or renames a tool.
• Test suite: 514 passing / 4 skipped, up from 478. Thirty-six net new tests, no assertion weakened or skipped.

See CHANGELOG.md for full details.