fix(hooks): a new session is shown the newest memories; the default briefing level is now minimal

[kevintseng]

Closes #401.

Two commits, each reviewed on its own:

1. fix(hooks): a new session is shown the newest memories, not the oldest, when scores tie
2. feat(briefing): the default briefing level is now minimal

What changes for a user

• A new session is shown the newest of a group of equally scored memories, not the oldest (SessionStart injects the oldest of a tied group: the hook's ranking has no recency for never-accessed entities and no tie-break #401). The SessionStart hook ranks a project's memories by confidence, use and recency and keeps the top few. Memories captured since the last daily decay carry one confidence value and, never accessed, score exactly alike; SQLite hands equal scores back in ascending id order, so the cut kept the oldest of them. The lesson query, whose pool is claimed first, had no ORDER BY at all and kept the five oldest lessons. Both now resolve newest first (id DESC as the last ORDER BY key of both scored forms, which also serves the global and recent pools that full adds; ORDER BY e.id DESC on the lesson query). A higher score still beats a newer memory.
• The default briefing level is now minimal (was standard; requested by the maintainer). A new session, and memesh briefing / the briefing tool, gets this project's decisions, lessons, known facts and recent activity, with the live repository state in front of them; the fresh task state and the durable-memory index come back with memesh config set briefing standard or MEMESH_BRIEFING=standard. An explicit setting is honoured exactly as before (env > config > default); an unset or invalid value now resolves to minimal. Only the constant moved; every statement of the old default (source comments, both MCP tool descriptions, AGENTS.md, three READMEs, the skill text, ARCHITECTURE, API_REFERENCE, the dashboard's Project tab in 11 locales, the doc-parity lock) was updated with it, and the tests that assumed the old default were made to say which level they need.

Evidence

• Real data first (issue SessionStart injects the oldest of a tied group: the hook's ranking has no recency for never-accessed entities and no tie-break #401): on a copy of a real graph the hook and memesh briefing disagreed at every level; the hook returned the oldest of 43 tied memories and the newest decision was number 43 of 43.
• npm run verify GREEN on the tree of each commit (commit 1: tree 66acd87a53a3; commit 2: tree 1ef0207d138f).
• Break-tests (string write-back, sha256 restored each time): removing only the lesson ordering turns 3 tests red; removing only the scored tie-breaks 5; all three 5; the two "a higher score still wins" controls stay green. Setting the default constant back to standard turns 36 tests in 4 files red.
• Reviews: commit 1 by four independent reviewers (fresh-eyes twice, code-reviewer, silent-failure-hunter); the first version had one BLOCKING finding (the unordered lesson query), fixed. Commit 2 by three independent reviewers (no blocking finding; the non-blocking ones are fixed in the commit or tracked in The effective briefing level is not shown anywhere, and no live journey runs at the default level #412).

Simplification pass

A read-only review of the whole diff looked for anything that could be cut without changing behaviour. Nothing was cut. It found two test-only candidates (about 11 lines) and no source or documentation candidate:

• tests/core/briefing.test.ts: the tie-parity test restores MEMESH_BRIEFING by hand, although the file's afterEach already calls vi.unstubAllEnvs(); vi.stubEnv would take two lines instead of nine.
• tests/hooks/session-start.test.ts: a five-line comment above the new describe block repeats the comment on the query it tests.

Neither changes behaviour, and applying them would restart the whole 14-check CI matrix (the Windows legs alone take about half an hour) for a cosmetic gain, so they are left. Kept on purpose, with the reason: the 12-line comment above the query in scripts/hooks/session-start.js (it records a measured SQLite behaviour that nothing else proves); the legacy-form leg of the parity test (nothing else compares the legacy hook's rendered text with the briefing's); the docs/ARCHITECTURE.md bullet (the only reader-facing statement of the tie rule); the two delete isolatedEnv.MEMESH_BRIEFING lines (they follow the existing per-helper pattern, and merging them would refactor code this PR does not own).

Not in this PR (tracked)

• Session-start pool selection: commits crowd out decisions, and the hook and the briefing use different windows and scores #411: the hook and the briefing still choose their pools differently (window sizes, candidate windows, the impact term), and a flood of auto-captured commits can still crowd a decision out of the project pool. This matters more now that minimal is the default. The proposed fix (select per layer in the shared leaf) changes what sessionLimit means and needs the owner's decision.
• MEMESH_RECIPIENT typos and path-shaped ids produce no signal, so the agent silently never hears of any message #402 to CLI inconsistencies: config get is missing, status says up to date on a newer version, briefing --index prints the framing #410: findings of an independent dogfood pass on 4.10.2.