Add additive SBD V6 metadata and timeline projection

[luochen-amd]

Summary

• add an additive SBD V6 projection while preserving the existing V5 schema version and execution behavior
• expose V6 metadata, outcome, timeline, and top-level close data
• record fail-open install and model_gate telemetry without changing preflight, gate, or resume outcomes
• implement the merged framework_agent timeline for configuration, source, Critic, macro-cycle, exit, and failure evidence
• project Framework-internal candidate discovery success, empty, and failure outcomes in execution order
• omit phase summaries and all token, pricing, and monetary cost fields

Review Fixes

• persist hard preflight failures even when they occur outside an instrumented install step
• acquire the session lock before writing resume-failure artifacts and fall back to an isolated failed-attempt session when the target session is busy
• preserve specialist source phases and exclude non-Framework specialist evidence from framework_agent timeline creation
• make legacy fixed-file timeline migration resumable after a partial migration

Compatibility

• keeps V5 output and state-machine behavior intact
• treats V6 collection as best-effort telemetry so collection failures do not affect optimization execution
• retains legacy Explore evidence only as input compatibility for the merged Framework Agent stage

Validation

• focused blocker reproductions: 4 passed
• SBD, Framework Agent, specialist, enablement, and attribution regression selection: 266 passed
• preflight regression selection: 237 passed
• Ruff checks passed for all changed Python files
• Ruff format checks passed for all changed Python files
• git diff --check passed

Environment Note

• Windows test runs inject a no-op fcntl compatibility module so POSIX-only KB imports can load; the production lock behavior remains covered by Linux CI and the focused busy-session test uses an explicit lock double

[github-actions]

CI E2E report — ❌ Timeout

item	value
result	❌ Timeout
model	Qwen/Qwen3-0.6B (dense)
resources	1× GPU, TP=1
PR branch	feat/sbd-v6-metadata-outcome-timeline
commit	4dd5533d8e43e5fcdfbf9c1b76458eb8ced759c7
session_id	52dcbb00-77e1-4fe7-8705-7dc00196e1fe
queue → dispatch	-68271s
run time	1163m 29s
total	25m 38s
reason	Timed out — the run never reached a terminal state in time (task stuck, or the GPU stayed queued too long).
detail	not terminal after 13200s

details

[luochen-amd]

Addressed the three blocking review findings:

• Persist hard preflight/install failures by creating a safe failure session, writing the install event, and materializing session_breakdown.json before re-raising the original exception.
• Preserve repeated install and model_gate executions in an ordered per-event timeline while keeping install.json and model_gate.json as latest-event compatibility projections. Legacy fixed-file sessions are migrated on the next write.
• Exclude Kernel-phase explore rebench operations from the Framework Agent timeline and require matching cycle/time window evidence.

Regression coverage now includes hard preflight failure persistence, fresh-to-resume event retention, legacy event migration, and Kernel Explore exclusion.

Local validation:

• SBD V6 focused tests: 26 passed
• Related exporter/model-gate suites: 215 passed, 1 Windows-only path assertion deselected
• Ruff check/format and git diff --check: passed

[luochen-amd]

This combined PR has been split to keep the review boundaries explicit:

• SBD V6 metadata, outcome, install, and model-gate timeline: Add SBD V6 metadata and startup timeline #1331 (based on main)
• Merged Framework Agent timeline and specialist phase attribution: feat/sbd-v6-framework-agent

The Framework branch is intentionally kept separate while #1301 and #1331 are still open. A focused Framework PR can be opened after the dependency bases are aligned, avoiding duplicate core changes in review.