2.0.8 batch: five Sentry-evidenced fixes (F-509, F-783, F-816, F-817, F-818)

[AminDhouib]

Stacked on #63 (it builds on the F-815 Sentry filter there). Retarget to main after #63 merges.

Five fixes, each driven by a live Sentry stacktrace, each TDD'd RED-first against the exact production error shape. Full hermetic lane: 1636 passed (+42 new tests), budget gate green, all LOC caps exact.

The fixes

F-817 — nodriver handler-table race (element_resolution.py) — Fixes STEALTH-CHROME-DEVTOOLS-MCP-2R, -2S
nodriver 0.47's remove_handler ends in a bare del self.handlers[evt_dom] that deletes the whole key; two overlapping Tab.wait()s race and the loser crashes with KeyError: <class 'nodriver.cdp.page.FrameStoppedLoading'>. The existing recovery wrapper now retries it like the -32000 stale-node case (one classifier, no parallel path); if it persists, a readable ToolError replaces the class-repr leak. Unrelated KeyErrors propagate untouched (two negative-control pins).

F-816 — CDP param-name folding (cdp_function_executor.py, 1010/1012) — Fixes -22
F-813 forgave command-name spelling; params still crashed on the CDP docs' own casing (awaitPromise). build_cdp_call folds incoming keys onto the resolved method's real signature; unknown params raise ToolError listing the valid names, through the method's KEEP contract.

F-818 — target-less browser + typed instance-not-found (browser_manager.py, exactly 1532/1532, line-neutral) — Fixes -2K, -12
Browser.get()'s bare next(filter(...)) turns into RuntimeError: coroutine raised StopIteration when no page target remains; the one exposed call site converts exactly that shape into an actionable ToolError. navigate's unknown-instance raise becomes InstanceNotFoundError, its timeout ToolError. Census of all 7 bare raises included in the commit; 4 left deliberately (F-811 operator-hint contract, internal invariants).

F-783 closed — error-convention sweep (dom_handler.py, server.py exactly 3411/3411) — Fixes -2H, -Y, -2F, -2P, -2V, -2Q, -2M, -2N
All 21 bare raise Exception sites in dom_handler + _with_cdp_timeout's timeout raise become ToolError (messages byte-identical). The F-815 filter now drops the family — which also dissolves the per-instance-UUID splitting of CDP-timeout issues. Deliberately kept shipping: cancellation (still CancelledError) and spawn failures (server.py:454 stays bare so the F-811 exhaustion class remains visible in Sentry). Audit finding doc annotated closed.

F-509 — reserved-port fallback (singleton.py exactly 1000/1000, proxy_forwarder.py) — Fixes -2J
Windows excluded-port ranges forbid binding 19222 (WinError 10013) but have no listener, so the connect-based foreign-held probe read the port as free and uvicorn died in the child on every start — seen on an external user's machine. New bind-and-close probe disqualifies a port only on PermissionError: treating any bind failure as disqualifying scattered the cold-start herd onto private ports (herd test 240s timeout vs 3.9s — the A/B that found it is pinned in TestOccupiedIsNotForbidden). Selection policy, restart_backend's contract, and identity/adoption gates unchanged.

Known residuals (flagged, not fixed here)

• F-509 A2 (DESIGN.md): a fleet cold-starting against a genuinely reserved port still scatters; needs port choice serialized under the cold-start lock.
• nodriver upstream: remove_handler's unguarded dels remain — we work around, not vendor-patch.
• dom_handler.py:859 ValueError for invalid scroll direction still ships to Sentry (chain contains a non-ToolError).
• DESIGN.md/RUNBOOK.md port-selection prose doesn't yet mention the permission dimension (no doc pin falsified).

Integration lane runs on this PR's CI (E2E across 3 OSes), same as #63.

[AminDhouib]

CI arbitration — the reds are pre-existing environmental findings, not this PR

Two runs, byte-identical trees (26fb2ec and empty-commit re-run 06356f7):

Cell	Run 1	Run 2	Verdict
integration (Windows/X64)	FAIL — pytest: null, no junit (the documented pre-2.0.3 herd-hang signature, evidence_F509_windows_herd_hang_2026-08-01.md)	PASS	flake, arbitrated green on an identical tree
integration (Linux/X64)	PASS	PASS	—
integration (macOS/ARM64)	FAIL — 3 nodes	FAIL — 2 nodes	F-806, see below
everything else (unit ×3, transport ×2, coverage, smoke ×6, aggregates' inputs)	PASS	PASS	—

The macOS failures are F-806 with new, conclusive evidence. Run 1's junit:

test_autodiscovery_and_cdp_match_image_chrome
AssertionError: ('Chrome/151.0.7922.76', '150.0.7871.187')  — assert '151' == '150'

The version probed early in the run was Chrome 150; the CDP-launched browser was Chrome 151 — Chrome's updater upgraded the binary mid-run on the macOS runner. That is exactly F-806's documented mechanism (pre-launch version probe memoized → masked UA contradicts the live browser's own client hints), which is why both stealth gates (ua_major_matches_client_hints, f770_product_masks_ua) fail in both runs. The "Event loop is closed" teardown noise is the separately-documented F-779.

This PR's diff has zero overlap with UA masking / platform_utils / Chrome discovery; any branch run on the current macOS image (including #63, if re-run today) inherits the same red until Chrome's updater and the image version converge — or until the preserved F-806 fix branch (fix/ua-version-skew @ b863878: memo re-keyed on binary identity + post-launch Browser.getVersion reconciliation) is finished and landed.

Local evidence for this PR: hermetic lane 1636 passed ×3 (twice pre-push hook, once direct), budget gate exit 0, all LOC caps exact.