fix(compression): apply the lossy policy to the default-combo fallback

[woodsonl]

Summary

• handleChatCore's legacy default-combo fallback now respects the x-omniroute-compression header and the fix(compression): keep lossy engines off the default path #14529 lossy policy: a plan the header chose (named combo, engine:, default, off) always wins, and on every other request the default combo runs only its safe steps. Lossy engines such as rtk stay off unless the request sends allow-lossy.
• planFromHeader now requires a non-empty array pipeline for a named-combo match. An empty-pipeline combo (reachable from the per-engine config page) used to win header precedence and let the builtin rtk/caveman pipeline run lossy on requests that never opted in. The same lookup resolved prototype keys (constructor, toString) to truthy garbage.
• The fallback logs a debug line when it skips for a request carrying a compression header, so operators can see why an edited default combo did not apply.

Related Issues

• Continues fix(compression): keep lossy engines off the default path #14529 (keep lossy engines off the default path).
• Preview parity contract from fix(backend): active compression profile shown in UI but response header reports compression off #12063.

Validation

• Change type: routing
• Focused tests and category gates from the golden path
• npm run lint (pre-commit: lint-staged, docs-sync, any-budget, tracked-artifacts all green)
• Reconciled with the current active release base; focused checks rerun afterward (upstream/release/v3.8.52 merged at 54e865d; 47/47 focused tests and typecheck:core green on the merged tree)
• Production-code changes include a new or updated automated test in this PR

⚠️ base-red inherited: #15306 (release/v3.8.52 tip was red when this PR opened; the failures are not from this branch)

Tests Added Or Updated

• tests/unit/compression/default-combo-lossy-policy.test.ts (new): the default combo's lossy steps are dropped without opt-in, kept under allow-lossy, and the header wins.
• tests/unit/chatcore-compression-combo-predicates.test.ts: eight new cases (yield-to-header, allow-lossy, empty pipeline, single lossy engine swapped for the safe pair, single safe engine kept, resolvable engine: header yields).
• tests/unit/compression/compression-header-dispatch.test.ts: two new cases (empty-pipeline combo is not a chosen plan, prototype-key header values resolve to null).
• tests/integration/chatcore-compression-integration.test.ts: removed the obsolete skipped case that expected rtk to run on the seeded default combo.

Coverage Notes

open-sse/ changes are covered by the three unit files above plus the integration file. The full compression aggregate (1682 tests) and typecheck:core pass on the merged tree. Evidence: _artifacts/suite-final.log and the QA report (verdict pass, 4/4 contracts) in .gstack/qa-reports/qa-report-compression-default-combo-2026-10-05.md.

Reviewer Notes

• Behavior change 1: a request with header "default" now gets the panel-derived default profile and no longer activates the DB default combo's outputMode or language packs. A client relying on that loses it silently; the response header still reports source=default.
• Behavior change 2: combos with an empty pipeline no longer win header precedence, so the builtin stacked pipeline no longer runs unrequested through that path.
• Known adjacent gaps left for follow-up on purpose (each needs its own TDD loop): adaptive floor-mode escalation can re-inject lossy engines after the policy (strategySelector.ts); the compressionPlan hand-off drops adaptive escalation when the default combo applies; routing-assigned compression combos still execute their raw pipeline while plan metadata reports the stripped one; the downgrade bypasses result memoization because session-dedup is not deterministic-eligible.