strata deposit: 16 entries — 2026-07-16 through 2026-07-18 backlog catch-up

[TapeshN]

Summary

Standard strata deposit pass — backlog catch-up. 16 undeposited LEARNINGS.md headers (roughly the last ~11 undeposited entries plus 5 new entries appended tonight, all dated 2026-07-16 through 2026-07-18) were reviewed, paraphrased IP-clean, deduplicated against the existing corpus, and staged as 16 individually-committed entries across guardrails/, infra/, and orchestration/.

• 16 entries staged (one commit per entry).
• 3 headers/sub-lessons skipped as near-duplicates of existing corpus entries (same generalizable pattern already deposited):
  • "re-review every fix on money/auth surfaces" — duplicates orchestration/2026-07-10-re-review-every-fix-on-money-and-auth-surfaces.md
  • "a new machine route unreachable until added to the routing allowlist" (live-payment webhook incident) — duplicates guardrails/2026-07-08-a-new-machine-route-can-be-silently-unreachable-until-added-to-the-allowlist.md + guardrails/2026-07-03-middleware-bypasses-handler-only-tests.md
  • "a chained-command gate matches heredoc prose as if it were the command" — duplicates guardrails/a-chained-command-gate-is-blind-to-what-changes-mid-chain.md (kept only the genuinely new half of that source entry: the ungated branch-rename surface)
• 1 header skipped entirely as a duplicate class: "worktree tests silently validated the PRIMARY's stale build" — same root mechanism (symlinked node_modules resolving into the primary checkout) as 3 existing entries (infra/2026-07-02-a-stale-workspace-symlink-can-silently-poison-the-primarys-build.md, infra/fresh-worktree-requires-package-install.md, infra/2026-07-07-a-worktree-only-test-failure-can-be-an-environment-artifact.md).
• Several dense source headers (bundling 3–7 sub-lessons each) were consolidated into fewer, richer strata entries rather than exploded 1:1, matching the corpus's existing "N lessons from one incident" precedent (e.g. guardrails/five-security-classes-build-tests-missed-adversarial-review-caught.md).
• Two Stripe/webhook credential-pairing incidents (one from a staging-setup session, one from a later dedicated incident) were merged into a single generalized entry since they're recurrences of the same class.
• All vendor/platform names in the bodies were genericized where it improved reusability (e.g. "a hosting platform" / "a database platform" instead of naming specific SaaS products) without losing the concrete mechanism.

IP / safety gates run

• TAP_CORE_BLOCKLIST verified present and readable before starting (fail-closed check).
• tools/strata/scrub.py run against all 16 paraphrased files individually — all CLEAN.
• tools/strata/scrub.py run against the raw LEARNINGS.md source text for these 16 headers — correctly flagged a LEAK hit (private-repo self-reference), confirming the paraphrase step actually stripped what needed stripping.
• Manual grep for residual PR numbers, task numbers, account-ID fragments, and domains across all 16 files — zero hits.
• .github/scripts/validate_entries.py run locally — all 16 new entries schema-valid, zero warnings attributable to them (451 entries valid org-wide; pre-existing warnings on other entries are unrelated).
• No floating docs — every new file lives inside an existing category directory.

Judge / LLM-as-judge status: the automated judge/verify-cert step in autofeed.py was not run for this batch (see floor note below) — treat this batch as judge-pending. Self-audit against proj-strata/.claude/agents/check.md's validation checklist was performed manually by the depositing agent instead (schema fields, IP gate, confidence-vs-body honesty, category sanity, no floating docs, deposit-mass flag acknowledged below).

Deposit mass: 16 entries in one PR exceeds the "focused deposits of ≤5 preferred" guidance (flagged, not blocking) — this is an explicit backlog catch-up pass across two nights of undeposited work, same as the precedent PR that shipped 19 entries in one batch.

ZERO ANTHROPIC_API_KEY USED. No autofeed.py --execute/--check/--dry-run was run with a key. All paraphrasing was done by the depositing agent's own inference (Claude Max plan). No metered API console was touched.

Test plan

• Operator reviews entries for accuracy, tone, and IP-cleanliness
• Operator merges when satisfied (draft — not auto-merge-enabled)

🤖 Generated with Claude Code