Extend --branch beyond context so a scoped push sends only the h5i
material linked to that branch, like git push:
  - notes: only provenance for commits reachable from the branch
  - objects: only manifests whose branch field matches
AST/msg/env/memory still push in full (AST is content-addressed and
deduplicated; the rest are genuinely cross-branch).

notes and objects are single aggregate refs shared by every branch, so
a scoped push must not force a filtered subset (that would delete other
branches' data on the remote). Instead it fetches the remote ref and
unions only this branch's entries onto it, then pushes a fast-forward
-- non-destructive to other branches.

New testable lib helpers:
  - repository::copy_scoped_notes_onto (overlays reachable-commit notes
    onto a remote-seeded dest ref; git2 handles notes fan-out)
  - objects::build_branch_scoped_merge (base manifests union local
    branch manifests, descends from base)

Tests: 6 lib unit tests (filtering + non-destructive base preservation +
empty cases) and 2 cli_integration tests (leak-prevention on a fresh
remote; non-destructive over a pre-seeded remote). MANUAL.md + man page
updated.

Previously `share push` pushed every branch's material unless --branch
was given. Flip it: with no flag the push now scopes to the current git
branch (context + notes + objects), matching git-push intuition. Add
--all-branches to opt back into the full push (first sync / CI), and make
it conflict with --branch.

Notes reachability falls back from refs/heads/<branch> to HEAD so a
detached checkout (common in CI) still scopes to its history instead of
pushing nothing.

Tests: replace push_without_branch_pushes_all_context_refs with
push_all_branches_pushes_every_context_ref; add push_defaults_to_current_branch
and push_branch_conflicts_with_all_branches. The two exact-OID pull
roundtrip tests now push --all-branches (the scoped default rebuilds a
per-branch union notes commit with a different OID by design).

Extend branch scoping past notes/objects to the two remaining
branch-linked refs:
  - msg: only messages tagged with the branch travel (sends auto-tag the
    current branch; replies inherit the thread's). The agent roster
    always travels in full.
  - env: only envs whose parent_branch matches travel — their
    manifests/policies/events (env meta) AND their code branches
    (refs/h5i/env/code/*). Their evidence captures (objects tagged with
    the env's own code branch) are carried too, by env_id, so the
    cross-clone review loop keeps its evidence.

memory stays full: it is a cumulative full-state snapshot chain whose
facts are branch-independent (kept-local on pull, no clean union).

A generic scoped_merge_push in the push handler drives the
non-destructive fetch+union+FF for objects/msg/env-meta. New lib
helpers: msg::build_branch_scoped_merge, env::build_branch_scoped_merge,
env::scoped_code_branch_refs, env::local_env_ids_for_branch (pub);
objects::build_branch_scoped_merge gains an also_env_ids param.

Tests: lib unit tests for msg/env scoping (filter + non-destructive base
+ empty) and objects env-evidence-by-env_id; a msg scoped-push
cli_integration test. The two cross-clone msg_integration sync tests now
push --all-branches (a review's branch tag differs from the git branch).
MANUAL.md + man page updated.

Drop the structural dimension: the ast.rs module (SemanticAst/AstDiff/
similarity/hashing), the refs/h5i/ast object store, semantic blame
(blame --mode ast), the AST structural diff command (h5i recall diff),
the --ast commit flag, and the python parser plugin + its discovery/
subprocess machinery. AST went unused and supporting all languages is
impractical.

Removed:
- src/ast.rs and plugin/h5i-py-parser.py (deleted)
- repository.rs: H5I_AST_REF, ast_parser plumbing in commit(), AST harvest,
  blame_by_ast, make_ast_parser, diff_ast, load_ast_at_commit,
  save_ast_to_sidecar, load_ast_blob, parser-script discovery + subprocess
  helpers, and their tests
- blame.rs: BlameMode (line-only now), the is_semantic* fields, dead
  BlameEntry/H5iBlameEntry structs
- metadata.rs: H5iCommitRecord.ast_hashes (old notes still deserialize —
  serde ignores unknown fields, verified by env_integration fixtures)
- error.rs: the Ast variant; main.rs: Commands::Diff, blame --mode, --ast,
  refs/h5i/ast in push/pull/patterns; mcp.rs blame mode; server.rs
  ast_file_count badge; storage.rs + cli_routing recall-diff
- docs: CLAUDE.md, MANUAL.md, man page, skill

blame stays (line-based, AI-provenance-annotated). Existing refs/h5i/ast
on clones/remotes are simply orphaned (harmless; no longer synced).

All suites green (lib 872, cli 79, env 101, msg 37, objects 25, +others);
clippy clean with and without --no-default-features.

Follow-up to the AST removal — update all docs/ pages and the MANUAL
source they derive from:
- MANUAL.md: drop AST from the share-push ref lists, the h5i_blame MCP
  mode param, and the (now-uncreated) ast/ init-dir note
- docs/manual/index.html: regenerated from MANUAL.md via scripts/gen_manual.py
- docs/index.html + docs/features/index.html: drop 'structural (AST)'
  from the dimensions tagline
- docs/guides/git-blame-for-ai-code: remove the AST-mode FAQ (incl. JSON-LD),
  the --mode ast example, the reformat-survival callout, and the stale
  ✨ semantic column in the sample output
- docs/blog/from-git-blame-to-ai-blame: remove the 'AST-mode blame' section

No AST/--mode ast/refs/h5i/ast references remain anywhere under docs/.