Session Log: D7 access defect filed, sized and dispositioned; Spike 5 opened

Date: 2026-08-24 (second session of the day; continues 2026-08-24-json-proxy-rescoped-and-endpoint-access-escalated.md) Participants: Than Grove, Claude Code Outcome: The 08-21 private-docs blocker cleared and the D7 access defect filed, verified, sized and dispositioned in one sitting. Spike 5 moved ○ Pending → ◐ Partial with 4 of 6 pass criteria resolved. Spike status key gained a missing definition.

Hand-written. The defect itself is not described here — it is tracked in uvalib/mandala-legacy-docs per the non-public documentation policy. Ask Than. Everything below is deliberately non-disclosing.

1. Private-docs access resolved — the 08-21 action for Yuji is closed

Than got access to both private repos (checked out at ~/Sandbox/Mandala/mandala-legacy-docs and ~/Sandbox/Mandala/mandala-navina-docs). The 08-21 ambiguity — repos missing, or merely invisible — resolves as invisible, not missing.

Filing destination per CONVENTION.md: routing keys on which stack the fix serves, not where the finding was made. This fix serves D7 → mandala-legacy-docs. Worth noting because the instinct is navina-docs: it was found during rebuild work and motivates a D11 requirement. The convention explicitly anticipates that case.

The 08-21 draft did not survive — both that day's transcripts were checked and contain only conversational prose. It was rewritten from Than's description and re-verified against production before filing, so it rests on fresh evidence rather than a three-day-old recollection. Public trail updated by PR #138, still saying only that a problem exists and who to ask.

2. The defect: verified, scoped, sized, decided — all in one session

Four rounds, each of which changed the answer:

Verified — reproduced anonymously against production, and found it broader than described.
Scoped — Than spot-checked the other three D7 apps; each restricts private nodes. Sources only, not platform-wide. Consistent with Spike 6's finding that the four apps diverged precisely in access enforcement.
Sized — ~2,200 published private records exposed. The structural detail worth keeping: the exposure is inherited, not explicit. Only 3 nodes carry an explicit private flag while 23,348 say "use group default"; privacy comes from 54 private collections. Sizing by searching for explicitly-private nodes would have found 3 and concluded it was trivial. Validated 10/10 on a live random sample.
Dispositioned — Than: do not patch D7. Severity Low. Fix it in D11. The records are bibliographic entries for published books available elsewhere; the unique content is the collection structure, also low value.

The technical arguments alone pointed toward patching — one site, small fix, access value already loaded, indefinite pre-cutover window. It was the content assessment that decided it. That distinction is written into the private record explicitly, because a future reader re-deriving the technical case would otherwise reach the opposite conclusion and wonder why nobody acted.

This makes the #136 agenda item the remediation, not just an architectural preference.

Two near-misses worth recording, both caught by checking rather than reasoning:

The system table put status in column 5, not where the pattern assumed — the Zotero modules read as disabled before the schema was confirmed. They are enabled.
The first sizing pass parsed only rows with quoted descriptions, silently dropping every built-in biblio type and reporting 36 types as "undefined".

3. Spike status key — "Complete" was undefined

Complete was used on Spike 4b but appeared nowhere in the key, so the one spike that closed on a failed hypothesis looked like a typo for Proven. Defined (PR #139):

Proven — the theory under test was demonstrated true.
Complete — question resolved, direction chosen, no blockers, but the original hypothesis did not pass. 4b's fail criterion fired; marking it Proven would imply the footnotes module worked, which is the opposite of what was found.

The key now also states what neither means: not "no work left". Spike 6 is Proven and still handed off its controllers. The status describes the state of the question, not the state of the code — which was the actual confusion.

4. Spike 5 opened — the main risk doesn't materialise

PR #140. ○ Pending → ◐ Partial.

bibcite is biblio's successor and ships biblio's own type list near-verbatim. The theory implicitly assumed a re-modelling onto CSL; it is a 1:1 name mapping. 26 of 36 types in use match, covering 94.2% of 25,629 rows.

Every gap is a Mandala custom type. Three are authorship distinctions D7 encoded as types (300 rows) which bibcite expresses through contributor roles — they collapse into book. Than confirmed the other four are wanted, with meanings (Review = book reviews; Block Print = a wood-block print of a Tibetan text). All four are config-only — a reference type is one YAML over a shared field pool, and bibcite's configurable type→CSL mapping has exact targets (review-book, entry-dictionary, article-newspaper, manuscript). The "critical reference type is missing" fail criterion resolves affirmatively rather than firing.

Than's semantics mattered directly: "Review" as book reviews is what makes review-book right rather than the weaker generic review.

Criterion 3 resolved by doing the work. Production runs biblio_style = "cse", which bibcite does not ship. Two importable configs prepared (CSE 9th and 8th, name-year) plus a D7 baseline, under docs/spikes/assets/spike-05-bibcite/ — deliberately not config/sync, since bibcite isn't installed and a cim would fail. Identifying the variant was the real work: the CSL repo carries seven CSE styles across three families, and D7's live output settles the family unambiguously while leaving the edition undetermined — so both are provided.

Also found: the citation route serves at least 8 styles, not one. harvard, vancouver and ieee would each need importing; deferred by Than until something actually requests them.

5. Recurring lesson, third instance today

The morning session recorded: for anything touching the edge/WAF, curl establishes nothing on its own. The afternoon added two more of the same shape — a claim that looks settled because nobody tested the other side of it:

	The untested assumption	What testing found
Morning	WAF blocks all Drupal hosts	Sources only
Afternoon	The access defect is platform-wide	Sources only
Afternoon	Sizing = count explicitly-private nodes	3 explicit vs ~2,200 inherited

The first two are the same site and the same inference error: generalising from one app to four when Spike 6 had already documented that the four diverged. Where D7 is concerned, per-site confirmation is the default, not the exception.

Artifacts


PR (merged)	#138 — private-repo access resolved; write-up filed
PR (open)	#139 — define "Complete" in the spike status key
PR (open)	#140 — Spike 5 desk research + CSE import job
Private	`mandala-legacy-docs` `d56a2da` → `a885e48` (scoped) → `2a84b13` (sized) → `2d3a1f9` (dispositioned)

Open items

#139 and #140 need review.
Group meeting: uniform endpoint access (#136) — Yuji + Xiaoming. Now also the remediation for the D7 defect.
Spike 5 next: is the single zotero_feed node still syncing (desk work, may remove criterion 4); then install bibcite in DDEV to verify the CSL mapping accepts custom type ids, and run criterion 6.
Never established: whether the D7 defect was ever exploited — needs D7 access logs for sources-api/*.
Unchanged from this morning: mandala-om #79 is merged but nothing was redeployed.